Skip to content

GitLab

Commit d1aac35d authored by zhangwenwei's avatar zhangwenwei
Browse files

Initial commit

parents
Hide whitespace changes
Inline Side-by-side
Showing with 3055 additions and 0 deletions
mmdet3d/core/evaluation/bbox_overlaps.py 0 → 100644
View file @ d1aac35d
import numpy as np
def bbox_overlaps(bboxes1, bboxes2, mode='iou'):
"""Calculate the ious between each bbox of bboxes1 and bboxes2.
Args:
bboxes1(ndarray): shape (n, 4)
bboxes2(ndarray): shape (k, 4)
mode(str): iou (intersection over union) or iof (intersection
over foreground)
Returns:
ious(ndarray): shape (n, k)
"""
assert mode in ['iou', 'iof']
bboxes1 = bboxes1.astype(np.float32)
bboxes2 = bboxes2.astype(np.float32)
rows = bboxes1.shape[0]
cols = bboxes2.shape[0]
ious = np.zeros((rows, cols), dtype=np.float32)
if rows * cols == 0:
return ious
exchange = False
if bboxes1.shape[0] > bboxes2.shape[0]:
bboxes1, bboxes2 = bboxes2, bboxes1
ious = np.zeros((cols, rows), dtype=np.float32)
exchange = True
area1 = (bboxes1[:, 2] - bboxes1[:, 0]) * (bboxes1[:, 3] - bboxes1[:, 1])
area2 = (bboxes2[:, 2] - bboxes2[:, 0]) * (bboxes2[:, 3] - bboxes2[:, 1])
for i in range(bboxes1.shape[0]):
x_start = np.maximum(bboxes1[i, 0], bboxes2[:, 0])
y_start = np.maximum(bboxes1[i, 1], bboxes2[:, 1])
x_end = np.minimum(bboxes1[i, 2], bboxes2[:, 2])
y_end = np.minimum(bboxes1[i, 3], bboxes2[:, 3])
overlap = np.maximum(x_end - x_start, 0) * np.maximum(
y_end - y_start, 0)
if mode == 'iou':
union = area1[i] + area2 - overlap
else:
union = area1[i] if not exchange else area2
ious[i, :] = overlap / union
if exchange:
ious = ious.T
return ious
mmdet3d/core/evaluation/class_names.py 0 → 100644
View file @ d1aac35d
import mmcv
def wider_face_classes():
return ['face']
def voc_classes():
return [
'aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat',
'chair', 'cow', 'diningtable', 'dog', 'horse', 'motorbike', 'person',
'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor'
]
def imagenet_det_classes():
return [
'accordion', 'airplane', 'ant', 'antelope', 'apple', 'armadillo',
'artichoke', 'axe', 'baby_bed', 'backpack', 'bagel', 'balance_beam',
'banana', 'band_aid', 'banjo', 'baseball', 'basketball', 'bathing_cap',
'beaker', 'bear', 'bee', 'bell_pepper', 'bench', 'bicycle', 'binder',
'bird', 'bookshelf', 'bow_tie', 'bow', 'bowl', 'brassiere', 'burrito',
'bus', 'butterfly', 'camel', 'can_opener', 'car', 'cart', 'cattle',
'cello', 'centipede', 'chain_saw', 'chair', 'chime', 'cocktail_shaker',
'coffee_maker', 'computer_keyboard', 'computer_mouse', 'corkscrew',
'cream', 'croquet_ball', 'crutch', 'cucumber', 'cup_or_mug', 'diaper',
'digital_clock', 'dishwasher', 'dog', 'domestic_cat', 'dragonfly',
'drum', 'dumbbell', 'electric_fan', 'elephant', 'face_powder', 'fig',
'filing_cabinet', 'flower_pot', 'flute', 'fox', 'french_horn', 'frog',
'frying_pan', 'giant_panda', 'goldfish', 'golf_ball', 'golfcart',
'guacamole', 'guitar', 'hair_dryer', 'hair_spray', 'hamburger',
'hammer', 'hamster', 'harmonica', 'harp', 'hat_with_a_wide_brim',
'head_cabbage', 'helmet', 'hippopotamus', 'horizontal_bar', 'horse',
'hotdog', 'iPod', 'isopod', 'jellyfish', 'koala_bear', 'ladle',
'ladybug', 'lamp', 'laptop', 'lemon', 'lion', 'lipstick', 'lizard',
'lobster', 'maillot', 'maraca', 'microphone', 'microwave', 'milk_can',
'miniskirt', 'monkey', 'motorcycle', 'mushroom', 'nail', 'neck_brace',
'oboe', 'orange', 'otter', 'pencil_box', 'pencil_sharpener', 'perfume',
'person', 'piano', 'pineapple', 'ping-pong_ball', 'pitcher', 'pizza',
'plastic_bag', 'plate_rack', 'pomegranate', 'popsicle', 'porcupine',
'power_drill', 'pretzel', 'printer', 'puck', 'punching_bag', 'purse',
'rabbit', 'racket', 'ray', 'red_panda', 'refrigerator',
'remote_control', 'rubber_eraser', 'rugby_ball', 'ruler',
'salt_or_pepper_shaker', 'saxophone', 'scorpion', 'screwdriver',
'seal', 'sheep', 'ski', 'skunk', 'snail', 'snake', 'snowmobile',
'snowplow', 'soap_dispenser', 'soccer_ball', 'sofa', 'spatula',
'squirrel', 'starfish', 'stethoscope', 'stove', 'strainer',
'strawberry', 'stretcher', 'sunglasses', 'swimming_trunks', 'swine',
'syringe', 'table', 'tape_player', 'tennis_ball', 'tick', 'tie',
'tiger', 'toaster', 'traffic_light', 'train', 'trombone', 'trumpet',
'turtle', 'tv_or_monitor', 'unicycle', 'vacuum', 'violin',
'volleyball', 'waffle_iron', 'washer', 'water_bottle', 'watercraft',
'whale', 'wine_bottle', 'zebra'
]
def imagenet_vid_classes():
return [
'airplane', 'antelope', 'bear', 'bicycle', 'bird', 'bus', 'car',
'cattle', 'dog', 'domestic_cat', 'elephant', 'fox', 'giant_panda',
'hamster', 'horse', 'lion', 'lizard', 'monkey', 'motorcycle', 'rabbit',
'red_panda', 'sheep', 'snake', 'squirrel', 'tiger', 'train', 'turtle',
'watercraft', 'whale', 'zebra'
]
def coco_classes():
return [
'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train',
'truck', 'boat', 'traffic_light', 'fire_hydrant', 'stop_sign',
'parking_meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep',
'cow', 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella',
'handbag', 'tie', 'suitcase', 'frisbee', 'skis', 'snowboard',
'sports_ball', 'kite', 'baseball_bat', 'baseball_glove', 'skateboard',
'surfboard', 'tennis_racket', 'bottle', 'wine_glass', 'cup', 'fork',
'knife', 'spoon', 'bowl', 'banana', 'apple', 'sandwich', 'orange',
'broccoli', 'carrot', 'hot_dog', 'pizza', 'donut', 'cake', 'chair',
'couch', 'potted_plant', 'bed', 'dining_table', 'toilet', 'tv',
'laptop', 'mouse', 'remote', 'keyboard', 'cell_phone', 'microwave',
'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase',
'scissors', 'teddy_bear', 'hair_drier', 'toothbrush'
]
def cityscapes_classes():
return [
'person', 'rider', 'car', 'truck', 'bus', 'train', 'motorcycle',
'bicycle'
]
def kitti_classes():
return [
'Car',
'Pedestrian',
'Cyclist',
'Van',
'Person_sitting',
]
dataset_aliases = {
'voc': ['voc', 'pascal_voc', 'voc07', 'voc12'],
'imagenet_det': ['det', 'imagenet_det', 'ilsvrc_det'],
'imagenet_vid': ['vid', 'imagenet_vid', 'ilsvrc_vid'],
'coco': ['coco', 'mscoco', 'ms_coco'],
'wider_face': ['WIDERFaceDataset', 'wider_face', 'WDIERFace'],
'cityscapes': ['cityscapes'],
'kitti': ['KITTI', 'kitti']
}
def get_classes(dataset):
"""Get class names of a dataset."""
alias2name = {}
for name, aliases in dataset_aliases.items():
for alias in aliases:
alias2name[alias] = name
if mmcv.is_str(dataset):
if dataset in alias2name:
labels = eval(alias2name[dataset] + '_classes()')
else:
raise ValueError('Unrecognized dataset: {}'.format(dataset))
else:
raise TypeError('dataset must a str, but got {}'.format(type(dataset)))
return labels
mmdet3d/core/evaluation/coco_utils.py 0 → 100644
View file @ d1aac35d
import itertools
import mmcv
import numpy as np
from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval
from terminaltables import AsciiTable
from .recall import eval_recalls
def coco_eval(result_files,
result_types,
coco,
max_dets=(100, 300, 1000),
cat_ids=[],
classwise=False):
for res_type in result_types:
assert res_type in [
'proposal', 'proposal_fast', 'bbox', 'segm', 'keypoints'
]
if mmcv.is_str(coco):
coco = COCO(coco)
assert isinstance(coco, COCO)
if result_types == ['proposal_fast']:
ar = fast_eval_recall(result_files, coco, np.array(max_dets))
for i, num in enumerate(max_dets):
print('AR@{}\t= {:.4f}'.format(num, ar[i]))
return
for res_type in result_types:
if isinstance(result_files, str):
result_file = result_files
elif isinstance(result_files, dict):
result_file = result_files[res_type]
else:
assert TypeError('result_files must be a str or dict')
assert result_file.endswith('.json')
coco_dets = coco.loadRes(result_file)
# it will load all images if cat_ids is []
# img_ids = getImgIds(coco, catIds=cat_ids)
if len(cat_ids) < 80:
img_ids = getImgIds(coco, catIds=cat_ids)
else:
img_ids = coco.getImgIds()
iou_type = 'bbox' if res_type == 'proposal' else res_type
cocoEval = COCOeval(coco, coco_dets, iou_type)
if cat_ids:
# cat_ids is not None means it is set
cocoEval.params.catIds = cat_ids
cocoEval.params.imgIds = img_ids
if res_type == 'proposal':
cocoEval.params.useCats = 0
cocoEval.params.maxDets = list(max_dets)
cocoEval.evaluate()
cocoEval.accumulate()
cocoEval.summarize()
if classwise:
# Compute per-category AP
# from https://github.com/facebookresearch/detectron2/blob/03064eb5bafe4a3e5750cc7a16672daf5afe8435/detectron2/evaluation/coco_evaluation.py#L259-L283 # noqa
precisions = cocoEval.eval['precision']
catIds = cat_ids if cat_ids else coco.getCatIds()
# precision has dims (iou, recall, cls, area range, max dets)
assert len(catIds) == precisions.shape[2]
results_per_category = []
for idx, catId in enumerate(catIds):
# area range index 0: all area ranges
# max dets index -1: typically 100 per image
nm = coco.loadCats(catId)[0]
precision = precisions[:, :, idx, 0, -1]
precision = precision[precision > -1]
ap = np.mean(precision) if precision.size else float('nan')
results_per_category.append(
('{}'.format(nm['name']),
'{:0.3f}'.format(float(ap * 100))))
N_COLS = min(6, len(results_per_category) * 2)
results_flatten = list(itertools.chain(*results_per_category))
headers = ['category', 'AP'] * (N_COLS // 2)
results_2d = itertools.zip_longest(
*[results_flatten[i::N_COLS] for i in range(N_COLS)])
table_data = [headers]
table_data += [result for result in results_2d]
table = AsciiTable(table_data)
print(table.table)
def fast_eval_recall(results,
coco,
max_dets,
iou_thrs=np.arange(0.5, 0.96, 0.05)):
if mmcv.is_str(results):
assert results.endswith('.pkl')
results = mmcv.load(results)
elif not isinstance(results, list):
raise TypeError(
'results must be a list of numpy arrays or a filename, not {}'.
format(type(results)))
gt_bboxes = []
img_ids = coco.getImgIds()
for i in range(len(img_ids)):
ann_ids = coco.getAnnIds(imgIds=img_ids[i])
ann_info = coco.loadAnns(ann_ids)
if len(ann_info) == 0:
gt_bboxes.append(np.zeros((0, 4)))
continue
bboxes = []
for ann in ann_info:
if ann.get('ignore', False) or ann['iscrowd']:
continue
x1, y1, w, h = ann['bbox']
bboxes.append([x1, y1, x1 + w, y1 + h])
bboxes = np.array(bboxes, dtype=np.float32)
if bboxes.shape[0] == 0:
bboxes = np.zeros((0, 4))
gt_bboxes.append(bboxes)
recalls = eval_recalls(
gt_bboxes, results, max_dets, iou_thrs, print_summary=False)
ar = recalls.mean(axis=1)
return ar
def xyxy2xywh(bbox):
_bbox = bbox.tolist()
return [
_bbox[0],
_bbox[1],
_bbox[2] - _bbox[0],
_bbox[3] - _bbox[1],
]
def proposal2json(dataset, results):
json_results = []
for idx in range(len(dataset)):
img_id = dataset.img_ids[idx]
bboxes = results[idx]
for i in range(bboxes.shape[0]):
data = dict()
data['image_id'] = img_id
data['bbox'] = xyxy2xywh(bboxes[i])
data['score'] = float(bboxes[i][4])
data['category_id'] = 1
json_results.append(data)
return json_results
def det2json(dataset, results):
json_results = []
for idx in range(len(dataset)):
img_id = dataset.img_ids[idx]
result = results[idx]
for label in range(len(result)):
bboxes = result[label]
for i in range(bboxes.shape[0]):
data = dict()
data['image_id'] = img_id
data['bbox'] = xyxy2xywh(bboxes[i])
data['score'] = float(bboxes[i][4])
data['category_id'] = dataset.cat_ids[label]
json_results.append(data)
return json_results
def segm2json(dataset, results):
bbox_json_results = []
segm_json_results = []
for idx in range(len(dataset)):
img_id = dataset.img_ids[idx]
det, seg = results[idx]
for label in range(len(det)):
# bbox results
bboxes = det[label]
for i in range(bboxes.shape[0]):
data = dict()
data['image_id'] = img_id
data['bbox'] = xyxy2xywh(bboxes[i])
data['score'] = float(bboxes[i][4])
data['category_id'] = dataset.cat_ids[label]
bbox_json_results.append(data)
# segm results
# some detectors use different score for det and segm
if isinstance(seg, tuple):
segms = seg[0][label]
mask_score = seg[1][label]
else:
segms = seg[label]
mask_score = [bbox[4] for bbox in bboxes]
for i in range(bboxes.shape[0]):
data = dict()
data['image_id'] = img_id
data['bbox'] = xyxy2xywh(bboxes[i])
data['score'] = float(mask_score[i])
data['category_id'] = dataset.cat_ids[label]
if isinstance(segms[i]['counts'], bytes):
segms[i]['counts'] = segms[i]['counts'].decode()
data['segmentation'] = segms[i]
segm_json_results.append(data)
return bbox_json_results, segm_json_results
def results2json(dataset, results, out_file):
result_files = dict()
if isinstance(results[0], list):
json_results = det2json(dataset, results)
result_files['bbox'] = '{}.{}.json'.format(out_file, 'bbox')
result_files['proposal'] = '{}.{}.json'.format(out_file, 'bbox')
mmcv.dump(json_results, result_files['bbox'])
elif isinstance(results[0], tuple):
json_results = segm2json(dataset, results)
result_files['bbox'] = '{}.{}.json'.format(out_file, 'bbox')
result_files['proposal'] = '{}.{}.json'.format(out_file, 'bbox')
result_files['segm'] = '{}.{}.json'.format(out_file, 'segm')
mmcv.dump(json_results[0], result_files['bbox'])
mmcv.dump(json_results[1], result_files['segm'])
elif isinstance(results[0], np.ndarray):
json_results = proposal2json(dataset, results)
result_files['proposal'] = '{}.{}.json'.format(out_file, 'proposal')
mmcv.dump(json_results, result_files['proposal'])
else:
raise TypeError('invalid type of results')
return result_files
def getImgIds(coco, imgIds=[], catIds=[]):
'''
Get img ids that satisfy given filter conditions.
Different from the coco.getImgIds, this function returns the id if
the img contains one of the cat rather than all.
:param imgIds (int array) : get imgs for given ids
:param catIds (int array) : get imgs with all given cats
:return: ids (int array) : integer array of img ids
'''
if len(imgIds) == len(catIds) == 0:
ids = coco.imgs.keys()
else:
ids = set(imgIds)
for i, catId in enumerate(catIds):
if i == 0 and len(ids) == 0:
ids = set(coco.catToImgs[catId])
else:
ids |= set(coco.catToImgs[catId])
return list(ids)
mmdet3d/core/evaluation/eval_hooks.py 0 → 100644
View file @ d1aac35d
import os
import os.path as osp
import mmcv
import numpy as np
import torch
import torch.distributed as dist
from mmcv.parallel import collate, scatter
from mmcv.runner import Hook
from pycocotools.cocoeval import COCOeval
from torch.utils.data import Dataset
from mmdet3d import datasets
from .coco_utils import fast_eval_recall, results2json
from .mean_ap import eval_map
class DistEvalHook(Hook):
def __init__(self, dataset, interval=1):
if isinstance(dataset, Dataset):
self.dataset = dataset
elif isinstance(dataset, dict):
self.dataset = datasets.build_dataset(dataset, {'test_mode': True})
else:
raise TypeError(
'dataset must be a Dataset object or a dict, not {}'.format(
type(dataset)))
self.interval = interval
def after_train_epoch(self, runner):
if not self.every_n_epochs(runner, self.interval):
return
runner.model.eval()
results = [None for _ in range(len(self.dataset))]
if runner.rank == 0:
prog_bar = mmcv.ProgressBar(len(self.dataset))
for idx in range(runner.rank, len(self.dataset), runner.world_size):
data = self.dataset[idx]
data_gpu = scatter(
collate([data], samples_per_gpu=1),
[torch.cuda.current_device()])[0]
# compute output
with torch.no_grad():
result = runner.model(
return_loss=False, rescale=True, **data_gpu)
results[idx] = result
batch_size = runner.world_size
if runner.rank == 0:
for _ in range(batch_size):
prog_bar.update()
if runner.rank == 0:
print('\n')
dist.barrier()
for i in range(1, runner.world_size):
tmp_file = osp.join(runner.work_dir, 'temp_{}.pkl'.format(i))
tmp_results = mmcv.load(tmp_file)
for idx in range(i, len(results), runner.world_size):
results[idx] = tmp_results[idx]
os.remove(tmp_file)
self.evaluate(runner, results)
else:
tmp_file = osp.join(runner.work_dir,
'temp_{}.pkl'.format(runner.rank))
mmcv.dump(results, tmp_file)
dist.barrier()
dist.barrier()
def evaluate(self):
raise NotImplementedError
class DistEvalmAPHook(DistEvalHook):
def evaluate(self, runner, results):
gt_bboxes = []
gt_labels = []
gt_ignore = []
for i in range(len(self.dataset)):
ann = self.dataset.get_ann_info(i)
bboxes = ann['bboxes']
labels = ann['labels']
if 'bboxes_ignore' in ann:
ignore = np.concatenate([
np.zeros(bboxes.shape[0], dtype=np.bool),
np.ones(ann['bboxes_ignore'].shape[0], dtype=np.bool)
])
gt_ignore.append(ignore)
bboxes = np.vstack([bboxes, ann['bboxes_ignore']])
labels = np.concatenate([labels, ann['labels_ignore']])
gt_bboxes.append(bboxes)
gt_labels.append(labels)
if not gt_ignore:
gt_ignore = None
# If the dataset is VOC2007, then use 11 points mAP evaluation.
if hasattr(self.dataset, 'year') and self.dataset.year == 2007:
ds_name = 'voc07'
else:
ds_name = self.dataset.CLASSES
mean_ap, eval_results = eval_map(
results,
gt_bboxes,
gt_labels,
gt_ignore=gt_ignore,
scale_ranges=None,
iou_thr=0.5,
dataset=ds_name,
print_summary=True)
runner.log_buffer.output['mAP'] = mean_ap
runner.log_buffer.ready = True
class KittiDistEvalmAPHook(DistEvalHook):
def evaluate(self, runner, results):
tmp_file = osp.join(runner.work_dir, 'temp_0')
if not isinstance(results[0], dict):
result_files = self.dataset.reformat_bbox(results, tmp_file)
paste_result, ret_dict = self.dataset.evaluate(result_files)
for ap_cls, ap_result in ret_dict.items():
for ap_type, ap in ap_result.items():
key = f'{ap_cls}_{ap_type}'
val = float('{:.4f}'.format(ap))
runner.log_buffer.output[key] = val
else:
for name in results[0]:
print('\nEvaluating {}'.format(name))
results_ = [out[name] for out in results]
tmp_file_ = osp.join(tmp_file, name)
result_files = self.dataset.reformat_bbox(results_, tmp_file_)
paste_result, ret_dict = self.dataset.evaluate(
result_files, name)
for ap_cls, ap_result in ret_dict.items():
for ap_type, ap in ap_result.items():
key = f'{name}/{ap_cls}_{ap_type}'
val = float('{:.4f}'.format(ap))
runner.log_buffer.output[key] = val
runner.log_buffer.ready = True
class CocoDistEvalRecallHook(DistEvalHook):
def __init__(self,
dataset,
interval=1,
proposal_nums=(100, 300, 1000),
iou_thrs=np.arange(0.5, 0.96, 0.05)):
super(CocoDistEvalRecallHook, self).__init__(
dataset, interval=interval)
self.proposal_nums = np.array(proposal_nums, dtype=np.int32)
self.iou_thrs = np.array(iou_thrs, dtype=np.float32)
def evaluate(self, runner, results):
# the official coco evaluation is too slow, here we use our own
# implementation instead, which may get slightly different results
ar = fast_eval_recall(results, self.dataset.coco, self.proposal_nums,
self.iou_thrs)
for i, num in enumerate(self.proposal_nums):
runner.log_buffer.output['AR@{}'.format(num)] = ar[i]
runner.log_buffer.ready = True
class CocoDistEvalmAPHook(DistEvalHook):
def evaluate(self, runner, results):
tmp_file = osp.join(runner.work_dir, 'temp_0')
result_files = results2json(self.dataset, results, tmp_file)
res_types = ['bbox', 'segm'
] if runner.model.module.with_mask else ['bbox']
cocoGt = self.dataset.coco
# load image based on cat_ids
if len(self.dataset.cat_ids) < len(self.dataset.CLASSES):
from .coco_utils import getImgIds
imgIds = getImgIds(cocoGt, catIds=self.dataset.cat_ids)
else:
imgIds = cocoGt.getImgIds()
for res_type in res_types:
try:
cocoDt = cocoGt.loadRes(result_files[res_type])
except IndexError:
print('No prediction found.')
break
iou_type = res_type
cocoEval = COCOeval(cocoGt, cocoDt, iou_type)
cocoEval.params.catIds = self.dataset.cat_ids
cocoEval.params.imgIds = imgIds
cocoEval.evaluate()
cocoEval.accumulate()
cocoEval.summarize()
metrics = ['mAP', 'mAP_50', 'mAP_75', 'mAP_s', 'mAP_m', 'mAP_l']
for i in range(len(metrics)):
key = '{}_{}'.format(res_type, metrics[i])
val = float('{:.3f}'.format(cocoEval.stats[i]))
runner.log_buffer.output[key] = val
runner.log_buffer.output['{}_mAP_copypaste'.format(res_type)] = (
'{ap[0]:.3f} {ap[1]:.3f} {ap[2]:.3f} {ap[3]:.3f} '
'{ap[4]:.3f} {ap[5]:.3f}').format(ap=cocoEval.stats[:6])
runner.log_buffer.ready = True
for res_type in res_types:
os.remove(result_files[res_type])
mmdet3d/core/evaluation/kitti_utils/__init__.py 0 → 100644
View file @ d1aac35d
from .eval import kitti_eval, kitti_eval_coco_style
__all__ = ['kitti_eval', 'kitti_eval_coco_style']
mmdet3d/core/evaluation/kitti_utils/eval.py 0 → 100644
View file @ d1aac35d
import gc
import io as sysio
import numba
import numpy as np
@numba.jit
def get_thresholds(scores: np.ndarray, num_gt, num_sample_pts=41):
scores.sort()
scores = scores[::-1]
current_recall = 0
thresholds = []
for i, score in enumerate(scores):
l_recall = (i + 1) / num_gt
if i < (len(scores) - 1):
r_recall = (i + 2) / num_gt
else:
r_recall = l_recall
if (((r_recall - current_recall) < (current_recall - l_recall))
and (i < (len(scores) - 1))):
continue
# recall = l_recall
thresholds.append(score)
current_recall += 1 / (num_sample_pts - 1.0)
return thresholds
def clean_data(gt_anno, dt_anno, current_class, difficulty):
CLASS_NAMES = ['car', 'pedestrian', 'cyclist']
MIN_HEIGHT = [40, 25, 25]
MAX_OCCLUSION = [0, 1, 2]
MAX_TRUNCATION = [0.15, 0.3, 0.5]
dc_bboxes, ignored_gt, ignored_dt = [], [], []
current_cls_name = CLASS_NAMES[current_class].lower()
num_gt = len(gt_anno['name'])
num_dt = len(dt_anno['name'])
num_valid_gt = 0
for i in range(num_gt):
bbox = gt_anno['bbox'][i]
gt_name = gt_anno['name'][i].lower()
height = bbox[3] - bbox[1]
valid_class = -1
if (gt_name == current_cls_name):
valid_class = 1
elif (current_cls_name == 'Pedestrian'.lower()
and 'Person_sitting'.lower() == gt_name):
valid_class = 0
elif (current_cls_name == 'Car'.lower() and 'Van'.lower() == gt_name):
valid_class = 0
else:
valid_class = -1
ignore = False
if ((gt_anno['occluded'][i] > MAX_OCCLUSION[difficulty])
or (gt_anno['truncated'][i] > MAX_TRUNCATION[difficulty])
or (height <= MIN_HEIGHT[difficulty])):
ignore = True
if valid_class == 1 and not ignore:
ignored_gt.append(0)
num_valid_gt += 1
elif (valid_class == 0 or (ignore and (valid_class == 1))):
ignored_gt.append(1)
else:
ignored_gt.append(-1)
# for i in range(num_gt):
if gt_anno['name'][i] == 'DontCare':
dc_bboxes.append(gt_anno['bbox'][i])
for i in range(num_dt):
if (dt_anno['name'][i].lower() == current_cls_name):
valid_class = 1
else:
valid_class = -1
height = abs(dt_anno['bbox'][i, 3] - dt_anno['bbox'][i, 1])
if height < MIN_HEIGHT[difficulty]:
ignored_dt.append(1)
elif valid_class == 1:
ignored_dt.append(0)
else:
ignored_dt.append(-1)
return num_valid_gt, ignored_gt, ignored_dt, dc_bboxes
@numba.jit(nopython=True)
def image_box_overlap(boxes, query_boxes, criterion=-1):
N = boxes.shape[0]
K = query_boxes.shape[0]
overlaps = np.zeros((N, K), dtype=boxes.dtype)
for k in range(K):
qbox_area = ((query_boxes[k, 2] - query_boxes[k, 0]) *
(query_boxes[k, 3] - query_boxes[k, 1]))
for n in range(N):
iw = (
min(boxes[n, 2], query_boxes[k, 2]) -
max(boxes[n, 0], query_boxes[k, 0]))
if iw > 0:
ih = (
min(boxes[n, 3], query_boxes[k, 3]) -
max(boxes[n, 1], query_boxes[k, 1]))
if ih > 0:
if criterion == -1:
ua = ((boxes[n, 2] - boxes[n, 0]) *
(boxes[n, 3] - boxes[n, 1]) + qbox_area -
iw * ih)
elif criterion == 0:
ua = ((boxes[n, 2] - boxes[n, 0]) *
(boxes[n, 3] - boxes[n, 1]))
elif criterion == 1:
ua = qbox_area
else:
ua = 1.0
overlaps[n, k] = iw * ih / ua
return overlaps
def bev_box_overlap(boxes, qboxes, criterion=-1):
from .rotate_iou import rotate_iou_gpu_eval
riou = rotate_iou_gpu_eval(boxes, qboxes, criterion)
return riou
@numba.jit(nopython=True, parallel=True)
def d3_box_overlap_kernel(boxes, qboxes, rinc, criterion=-1):
# ONLY support overlap in CAMERA, not lidar.
# TODO: change to use prange for parallel mode, should check the difference
N, K = boxes.shape[0], qboxes.shape[0]
for i in numba.prange(N):
for j in numba.prange(K):
if rinc[i, j] > 0:
# iw = (min(boxes[i, 1] + boxes[i, 4], qboxes[j, 1] +
# qboxes[j, 4]) - max(boxes[i, 1], qboxes[j, 1]))
iw = (
min(boxes[i, 1], qboxes[j, 1]) -
max(boxes[i, 1] - boxes[i, 4],
qboxes[j, 1] - qboxes[j, 4]))
if iw > 0:
area1 = boxes[i, 3] * boxes[i, 4] * boxes[i, 5]
area2 = qboxes[j, 3] * qboxes[j, 4] * qboxes[j, 5]
inc = iw * rinc[i, j]
if criterion == -1:
ua = (area1 + area2 - inc)
elif criterion == 0:
ua = area1
elif criterion == 1:
ua = area2
else:
ua = inc
rinc[i, j] = inc / ua
else:
rinc[i, j] = 0.0
def d3_box_overlap(boxes, qboxes, criterion=-1):
from .rotate_iou import rotate_iou_gpu_eval
rinc = rotate_iou_gpu_eval(boxes[:, [0, 2, 3, 5, 6]],
qboxes[:, [0, 2, 3, 5, 6]], 2)
d3_box_overlap_kernel(boxes, qboxes, rinc, criterion)
return rinc
@numba.jit(nopython=True)
def compute_statistics_jit(overlaps,
gt_datas,
dt_datas,
ignored_gt,
ignored_det,
dc_bboxes,
metric,
min_overlap,
thresh=0,
compute_fp=False,
compute_aos=False):
det_size = dt_datas.shape[0]
gt_size = gt_datas.shape[0]
dt_scores = dt_datas[:, -1]
dt_alphas = dt_datas[:, 4]
gt_alphas = gt_datas[:, 4]
dt_bboxes = dt_datas[:, :4]
# gt_bboxes = gt_datas[:, :4]
assigned_detection = [False] * det_size
ignored_threshold = [False] * det_size
if compute_fp:
for i in range(det_size):
if (dt_scores[i] < thresh):
ignored_threshold[i] = True
NO_DETECTION = -10000000
tp, fp, fn, similarity = 0, 0, 0, 0
# thresholds = [0.0]
# delta = [0.0]
thresholds = np.zeros((gt_size, ))
thresh_idx = 0
delta = np.zeros((gt_size, ))
delta_idx = 0
for i in range(gt_size):
if ignored_gt[i] == -1:
continue
det_idx = -1
valid_detection = NO_DETECTION
max_overlap = 0
assigned_ignored_det = False
for j in range(det_size):
if (ignored_det[j] == -1):
continue
if (assigned_detection[j]):
continue
if (ignored_threshold[j]):
continue
overlap = overlaps[j, i]
dt_score = dt_scores[j]
if (not compute_fp and (overlap > min_overlap)
and dt_score > valid_detection):
det_idx = j
valid_detection = dt_score
elif (compute_fp and (overlap > min_overlap)
and (overlap > max_overlap or assigned_ignored_det)
and ignored_det[j] == 0):
max_overlap = overlap
det_idx = j
valid_detection = 1
assigned_ignored_det = False
elif (compute_fp and (overlap > min_overlap)
and (valid_detection == NO_DETECTION)
and ignored_det[j] == 1):
det_idx = j
valid_detection = 1
assigned_ignored_det = True
if (valid_detection == NO_DETECTION) and ignored_gt[i] == 0:
fn += 1
elif ((valid_detection != NO_DETECTION)
and (ignored_gt[i] == 1 or ignored_det[det_idx] == 1)):
assigned_detection[det_idx] = True
elif valid_detection != NO_DETECTION:
tp += 1
# thresholds.append(dt_scores[det_idx])
thresholds[thresh_idx] = dt_scores[det_idx]
thresh_idx += 1
if compute_aos:
# delta.append(gt_alphas[i] - dt_alphas[det_idx])
delta[delta_idx] = gt_alphas[i] - dt_alphas[det_idx]
delta_idx += 1
assigned_detection[det_idx] = True
if compute_fp:
for i in range(det_size):
if (not (assigned_detection[i] or ignored_det[i] == -1
or ignored_det[i] == 1 or ignored_threshold[i])):
fp += 1
nstuff = 0
if metric == 0:
overlaps_dt_dc = image_box_overlap(dt_bboxes, dc_bboxes, 0)
for i in range(dc_bboxes.shape[0]):
for j in range(det_size):
if (assigned_detection[j]):
continue
if (ignored_det[j] == -1 or ignored_det[j] == 1):
continue
if (ignored_threshold[j]):
continue
if overlaps_dt_dc[j, i] > min_overlap:
assigned_detection[j] = True
nstuff += 1
fp -= nstuff
if compute_aos:
tmp = np.zeros((fp + delta_idx, ))
# tmp = [0] * fp
for i in range(delta_idx):
tmp[i + fp] = (1.0 + np.cos(delta[i])) / 2.0
# tmp.append((1.0 + np.cos(delta[i])) / 2.0)
# assert len(tmp) == fp + tp
# assert len(delta) == tp
if tp > 0 or fp > 0:
similarity = np.sum(tmp)
else:
similarity = -1
return tp, fp, fn, similarity, thresholds[:thresh_idx]
def get_split_parts(num, num_part):
same_part = num // num_part
remain_num = num % num_part
if remain_num == 0:
return [same_part] * num_part
else:
return [same_part] * num_part + [remain_num]
@numba.jit(nopython=True)
def fused_compute_statistics(overlaps,
pr,
gt_nums,
dt_nums,
dc_nums,
gt_datas,
dt_datas,
dontcares,
ignored_gts,
ignored_dets,
metric,
min_overlap,
thresholds,
compute_aos=False):
gt_num = 0
dt_num = 0
dc_num = 0
for i in range(gt_nums.shape[0]):
for t, thresh in enumerate(thresholds):
overlap = overlaps[dt_num:dt_num + dt_nums[i],
gt_num:gt_num + gt_nums[i]]
gt_data = gt_datas[gt_num:gt_num + gt_nums[i]]
dt_data = dt_datas[dt_num:dt_num + dt_nums[i]]
ignored_gt = ignored_gts[gt_num:gt_num + gt_nums[i]]
ignored_det = ignored_dets[dt_num:dt_num + dt_nums[i]]
dontcare = dontcares[dc_num:dc_num + dc_nums[i]]
tp, fp, fn, similarity, _ = compute_statistics_jit(
overlap,
gt_data,
dt_data,
ignored_gt,
ignored_det,
dontcare,
metric,
min_overlap=min_overlap,
thresh=thresh,
compute_fp=True,
compute_aos=compute_aos)
pr[t, 0] += tp
pr[t, 1] += fp
pr[t, 2] += fn
if similarity != -1:
pr[t, 3] += similarity
gt_num += gt_nums[i]
dt_num += dt_nums[i]
dc_num += dc_nums[i]
def calculate_iou_partly(gt_annos, dt_annos, metric, num_parts=50):
"""fast iou algorithm. this function can be used independently to
do result analysis. Must be used in CAMERA coordinate system.
Args:
gt_annos: dict, must from get_label_annos() in kitti_common.py
dt_annos: dict, must from get_label_annos() in kitti_common.py
metric: eval type. 0: bbox, 1: bev, 2: 3d
num_parts: int. a parameter for fast calculate algorithm
"""
assert len(gt_annos) == len(dt_annos)
total_dt_num = np.stack([len(a['name']) for a in dt_annos], 0)
total_gt_num = np.stack([len(a['name']) for a in gt_annos], 0)
num_examples = len(gt_annos)
split_parts = get_split_parts(num_examples, num_parts)
parted_overlaps = []
example_idx = 0
for num_part in split_parts:
gt_annos_part = gt_annos[example_idx:example_idx + num_part]
dt_annos_part = dt_annos[example_idx:example_idx + num_part]
if metric == 0:
gt_boxes = np.concatenate([a['bbox'] for a in gt_annos_part], 0)
dt_boxes = np.concatenate([a['bbox'] for a in dt_annos_part], 0)
overlap_part = image_box_overlap(gt_boxes, dt_boxes)
elif metric == 1:
loc = np.concatenate(
[a['location'][:, [0, 2]] for a in gt_annos_part], 0)
dims = np.concatenate(
[a['dimensions'][:, [0, 2]] for a in gt_annos_part], 0)
rots = np.concatenate([a['rotation_y'] for a in gt_annos_part], 0)
gt_boxes = np.concatenate([loc, dims, rots[..., np.newaxis]],
axis=1)
loc = np.concatenate(
[a['location'][:, [0, 2]] for a in dt_annos_part], 0)
dims = np.concatenate(
[a['dimensions'][:, [0, 2]] for a in dt_annos_part], 0)
rots = np.concatenate([a['rotation_y'] for a in dt_annos_part], 0)
dt_boxes = np.concatenate([loc, dims, rots[..., np.newaxis]],
axis=1)
overlap_part = bev_box_overlap(gt_boxes,
dt_boxes).astype(np.float64)
elif metric == 2:
loc = np.concatenate([a['location'] for a in gt_annos_part], 0)
dims = np.concatenate([a['dimensions'] for a in gt_annos_part], 0)
rots = np.concatenate([a['rotation_y'] for a in gt_annos_part], 0)
gt_boxes = np.concatenate([loc, dims, rots[..., np.newaxis]],
axis=1)
loc = np.concatenate([a['location'] for a in dt_annos_part], 0)
dims = np.concatenate([a['dimensions'] for a in dt_annos_part], 0)
rots = np.concatenate([a['rotation_y'] for a in dt_annos_part], 0)
dt_boxes = np.concatenate([loc, dims, rots[..., np.newaxis]],
axis=1)
overlap_part = d3_box_overlap(gt_boxes,
dt_boxes).astype(np.float64)
else:
raise ValueError('unknown metric')
parted_overlaps.append(overlap_part)
example_idx += num_part
overlaps = []
example_idx = 0
for j, num_part in enumerate(split_parts):
gt_annos_part = gt_annos[example_idx:example_idx + num_part]
dt_annos_part = dt_annos[example_idx:example_idx + num_part]
gt_num_idx, dt_num_idx = 0, 0
for i in range(num_part):
gt_box_num = total_gt_num[example_idx + i]
dt_box_num = total_dt_num[example_idx + i]
overlaps.append(
parted_overlaps[j][gt_num_idx:gt_num_idx + gt_box_num,
dt_num_idx:dt_num_idx + dt_box_num])
gt_num_idx += gt_box_num
dt_num_idx += dt_box_num
example_idx += num_part
return overlaps, parted_overlaps, total_gt_num, total_dt_num
def _prepare_data(gt_annos, dt_annos, current_class, difficulty):
gt_datas_list = []
dt_datas_list = []
total_dc_num = []
ignored_gts, ignored_dets, dontcares = [], [], []
total_num_valid_gt = 0
for i in range(len(gt_annos)):
rets = clean_data(gt_annos[i], dt_annos[i], current_class, difficulty)
num_valid_gt, ignored_gt, ignored_det, dc_bboxes = rets
ignored_gts.append(np.array(ignored_gt, dtype=np.int64))
ignored_dets.append(np.array(ignored_det, dtype=np.int64))
if len(dc_bboxes) == 0:
dc_bboxes = np.zeros((0, 4)).astype(np.float64)
else:
dc_bboxes = np.stack(dc_bboxes, 0).astype(np.float64)
total_dc_num.append(dc_bboxes.shape[0])
dontcares.append(dc_bboxes)
total_num_valid_gt += num_valid_gt
gt_datas = np.concatenate(
[gt_annos[i]['bbox'], gt_annos[i]['alpha'][..., np.newaxis]], 1)
dt_datas = np.concatenate([
dt_annos[i]['bbox'], dt_annos[i]['alpha'][..., np.newaxis],
dt_annos[i]['score'][..., np.newaxis]
], 1)
gt_datas_list.append(gt_datas)
dt_datas_list.append(dt_datas)
total_dc_num = np.stack(total_dc_num, axis=0)
return (gt_datas_list, dt_datas_list, ignored_gts, ignored_dets, dontcares,
total_dc_num, total_num_valid_gt)
def eval_class(gt_annos,
dt_annos,
current_classes,
difficultys,
metric,
min_overlaps,
compute_aos=False,
num_parts=200):
"""Kitti eval. support 2d/bev/3d/aos eval. support 0.5:0.05:0.95 coco AP.
Args:
gt_annos: dict, must from get_label_annos() in kitti_common.py
dt_annos: dict, must from get_label_annos() in kitti_common.py
current_classes: list of int, 0: car, 1: pedestrian, 2: cyclist
difficultys: list of int. eval difficulty, 0: easy, 1: normal, 2: hard
metric: eval type. 0: bbox, 1: bev, 2: 3d
min_overlaps: float, min overlap. format: [num_overlap, metric, class].
num_parts: int. a parameter for fast calculate algorithm
Returns:
dict of recall, precision and aos
"""
assert len(gt_annos) == len(dt_annos)
num_examples = len(gt_annos)
split_parts = get_split_parts(num_examples, num_parts)
rets = calculate_iou_partly(dt_annos, gt_annos, metric, num_parts)
overlaps, parted_overlaps, total_dt_num, total_gt_num = rets
N_SAMPLE_PTS = 41
num_minoverlap = len(min_overlaps)
num_class = len(current_classes)
num_difficulty = len(difficultys)
precision = np.zeros(
[num_class, num_difficulty, num_minoverlap, N_SAMPLE_PTS])
recall = np.zeros(
[num_class, num_difficulty, num_minoverlap, N_SAMPLE_PTS])
aos = np.zeros([num_class, num_difficulty, num_minoverlap, N_SAMPLE_PTS])
for m, current_class in enumerate(current_classes):
for l, difficulty in enumerate(difficultys):
rets = _prepare_data(gt_annos, dt_annos, current_class, difficulty)
(gt_datas_list, dt_datas_list, ignored_gts, ignored_dets,
dontcares, total_dc_num, total_num_valid_gt) = rets
for k, min_overlap in enumerate(min_overlaps[:, metric, m]):
thresholdss = []
for i in range(len(gt_annos)):
rets = compute_statistics_jit(
overlaps[i],
gt_datas_list[i],
dt_datas_list[i],
ignored_gts[i],
ignored_dets[i],
dontcares[i],
metric,
min_overlap=min_overlap,
thresh=0.0,
compute_fp=False)
tp, fp, fn, similarity, thresholds = rets
thresholdss += thresholds.tolist()
thresholdss = np.array(thresholdss)
thresholds = get_thresholds(thresholdss, total_num_valid_gt)
thresholds = np.array(thresholds)
pr = np.zeros([len(thresholds), 4])
idx = 0
for j, num_part in enumerate(split_parts):
gt_datas_part = np.concatenate(
gt_datas_list[idx:idx + num_part], 0)
dt_datas_part = np.concatenate(
dt_datas_list[idx:idx + num_part], 0)
dc_datas_part = np.concatenate(
dontcares[idx:idx + num_part], 0)
ignored_dets_part = np.concatenate(
ignored_dets[idx:idx + num_part], 0)
ignored_gts_part = np.concatenate(
ignored_gts[idx:idx + num_part], 0)
fused_compute_statistics(
parted_overlaps[j],
pr,
total_gt_num[idx:idx + num_part],
total_dt_num[idx:idx + num_part],
total_dc_num[idx:idx + num_part],
gt_datas_part,
dt_datas_part,
dc_datas_part,
ignored_gts_part,
ignored_dets_part,
metric,
min_overlap=min_overlap,
thresholds=thresholds,
compute_aos=compute_aos)
idx += num_part
for i in range(len(thresholds)):
recall[m, l, k, i] = pr[i, 0] / (pr[i, 0] + pr[i, 2])
precision[m, l, k, i] = pr[i, 0] / (pr[i, 0] + pr[i, 1])
if compute_aos:
aos[m, l, k, i] = pr[i, 3] / (pr[i, 0] + pr[i, 1])
for i in range(len(thresholds)):
precision[m, l, k, i] = np.max(
precision[m, l, k, i:], axis=-1)
recall[m, l, k, i] = np.max(recall[m, l, k, i:], axis=-1)
if compute_aos:
aos[m, l, k, i] = np.max(aos[m, l, k, i:], axis=-1)
ret_dict = {
'recall': recall,
'precision': precision,
'orientation': aos,
}
# clean temp variables
del overlaps
del parted_overlaps
gc.collect()
return ret_dict
def get_mAP(prec):
sums = 0
for i in range(0, prec.shape[-1], 4):
sums = sums + prec[..., i]
return sums / 11 * 100
def print_str(value, *arg, sstream=None):
if sstream is None:
sstream = sysio.StringIO()
sstream.truncate(0)
sstream.seek(0)
print(value, *arg, file=sstream)
return sstream.getvalue()
def do_eval(gt_annos,
dt_annos,
current_classes,
min_overlaps,
eval_types=['bbox', 'bev', '3d']):
# min_overlaps: [num_minoverlap, metric, num_class]
difficultys = [0, 1, 2]
ret = eval_class(
gt_annos,
dt_annos,
current_classes,
difficultys,
0,
min_overlaps,
compute_aos=('aos' in eval_types))
# ret: [num_class, num_diff, num_minoverlap, num_sample_points]
mAP_bbox = get_mAP(ret['precision'])
mAP_aos = None
if 'aos' in eval_types:
mAP_aos = get_mAP(ret['orientation'])
mAP_bev = None
if 'bev' in eval_types:
ret = eval_class(gt_annos, dt_annos, current_classes, difficultys, 1,
min_overlaps)
mAP_bev = get_mAP(ret['precision'])
mAP_3d = None
if '3d' in eval_types:
ret = eval_class(gt_annos, dt_annos, current_classes, difficultys, 2,
min_overlaps)
mAP_3d = get_mAP(ret['precision'])
return mAP_bbox, mAP_bev, mAP_3d, mAP_aos
def do_coco_style_eval(gt_annos, dt_annos, current_classes, overlap_ranges,
compute_aos):
# overlap_ranges: [range, metric, num_class]
min_overlaps = np.zeros([10, *overlap_ranges.shape[1:]])
for i in range(overlap_ranges.shape[1]):
for j in range(overlap_ranges.shape[2]):
min_overlaps[:, i, j] = np.linspace(*overlap_ranges[:, i, j])
mAP_bbox, mAP_bev, mAP_3d, mAP_aos = do_eval(gt_annos, dt_annos,
current_classes, min_overlaps,
compute_aos)
# ret: [num_class, num_diff, num_minoverlap]
mAP_bbox = mAP_bbox.mean(-1)
mAP_bev = mAP_bev.mean(-1)
mAP_3d = mAP_3d.mean(-1)
if mAP_aos is not None:
mAP_aos = mAP_aos.mean(-1)
return mAP_bbox, mAP_bev, mAP_3d, mAP_aos
def kitti_eval(gt_annos,
dt_annos,
current_classes,
eval_types=['bbox', 'bev', '3d']):
assert 'bbox' in eval_types, 'must evaluate bbox at least'
overlap_0_7 = np.array([[0.7, 0.5, 0.5, 0.7,
0.5], [0.7, 0.5, 0.5, 0.7, 0.5],
[0.7, 0.5, 0.5, 0.7, 0.5]])
overlap_0_5 = np.array([[0.7, 0.5, 0.5, 0.7, 0.5],
[0.5, 0.25, 0.25, 0.5, 0.25],
[0.5, 0.25, 0.25, 0.5, 0.25]])
min_overlaps = np.stack([overlap_0_7, overlap_0_5], axis=0) # [2, 3, 5]
class_to_name = {
0: 'Car',
1: 'Pedestrian',
2: 'Cyclist',
3: 'Van',
4: 'Person_sitting',
}
name_to_class = {v: n for n, v in class_to_name.items()}
if not isinstance(current_classes, (list, tuple)):
current_classes = [current_classes]
current_classes_int = []
for curcls in current_classes:
if isinstance(curcls, str):
current_classes_int.append(name_to_class[curcls])
else:
current_classes_int.append(curcls)
current_classes = current_classes_int
min_overlaps = min_overlaps[:, :, current_classes]
result = ''
# check whether alpha is valid
compute_aos = False
for anno in dt_annos:
if anno['alpha'].shape[0] != 0:
if anno['alpha'][0] != -10:
compute_aos = True
eval_types.append('aos')
break
mAPbbox, mAPbev, mAP3d, mAPaos = do_eval(gt_annos, dt_annos,
current_classes, min_overlaps,
eval_types)
ret_dict = {}
difficulty = ['easy', 'moderate', 'hard']
for j, curcls in enumerate(current_classes):
# mAP threshold array: [num_minoverlap, metric, class]
# mAP result: [num_class, num_diff, num_minoverlap]
curcls_name = class_to_name[curcls]
ret_dict[curcls_name] = {}
for i in range(min_overlaps.shape[0]):
# prepare results for print
result += ('{} AP@{:.2f}, {:.2f}, {:.2f}:\n'.format(
curcls_name, *min_overlaps[i, :, j]))
if mAPbbox is not None:
result += (
'bbox AP:{:.4f}, {:.4f}, {:.4f}\n'.format(*mAPbbox[j, :,
i]))
if mAPbev is not None:
result += (
'bev AP:{:.4f}, {:.4f}, {:.4f}\n'.format(*mAPbev[j, :,
i]))
if mAP3d is not None:
result += (
'3d AP:{:.4f}, {:.4f}, {:.4f}\n'.format(*mAP3d[j, :, i]))
if compute_aos:
result += (
'aos AP:{:.2f}, {:.2f}, {:.2f}\n'.format(*mAPaos[j, :,
i]))
# prepare results for logger
for idx in range(3):
postfix = '{}_{}'.format(difficulty[idx], min_overlaps[i, idx,
j])
if mAP3d is not None:
ret_dict[curcls_name]['3D_{}'.format(postfix)] = mAP3d[j,
idx,
i]
if mAPbev is not None:
ret_dict[curcls_name]['BEV_{}'.format(postfix)] = mAPbev[
j, idx, i]
if mAPbbox is not None:
ret_dict[curcls_name]['2D_{}'.format(postfix)] = mAPbbox[
j, idx, i]
# calculate mAP over all classes if there are multiple classes
if len(current_classes) > 1:
# prepare results for print
result += ('\nOverall AP@{}, {}, {}:\n'.format(*difficulty))
if mAPbbox is not None:
mAPbbox = mAPbbox.mean(axis=0)
result += ('bbox AP:{:.4f}, {:.4f}, {:.4f}\n'.format(*mAPbbox[:,
0]))
if mAPbev is not None:
mAPbev = mAPbev.mean(axis=0)
result += ('bev AP:{:.4f}, {:.4f}, {:.4f}\n'.format(*mAPbev[:,
0]))
if mAP3d is not None:
mAP3d = mAP3d.mean(axis=0)
result += ('3d AP:{:.4f}, {:.4f}, {:.4f}\n'.format(*mAP3d[:, 0]))
if compute_aos:
mAPaos = mAPaos.mean(axis=0)
result += ('aos AP:{:.2f}, {:.2f}, {:.2f}\n'.format(*mAPaos[:,
0]))
# prepare results for logger
ret_dict['Overall'] = dict()
for idx in range(3):
postfix = '{}'.format(difficulty[idx])
if mAP3d is not None:
ret_dict['Overall']['3D_{}'.format(postfix)] = mAP3d[idx, 0]
if mAPbev is not None:
ret_dict['Overall']['BEV_{}'.format(postfix)] = mAPbev[idx, 0]
if mAPbbox is not None:
ret_dict['Overall']['2D_{}'.format(postfix)] = mAPbbox[idx, 0]
print(result)
return result, ret_dict
def kitti_eval_coco_style(gt_annos, dt_annos, current_classes):
class_to_name = {
0: 'Car',
1: 'Pedestrian',
2: 'Cyclist',
3: 'Van',
4: 'Person_sitting',
}
class_to_range = {
0: [0.5, 0.95, 10],
1: [0.25, 0.7, 10],
2: [0.25, 0.7, 10],
3: [0.5, 0.95, 10],
4: [0.25, 0.7, 10],
}
name_to_class = {v: n for n, v in class_to_name.items()}
if not isinstance(current_classes, (list, tuple)):
current_classes = [current_classes]
current_classes_int = []
for curcls in current_classes:
if isinstance(curcls, str):
current_classes_int.append(name_to_class[curcls])
else:
current_classes_int.append(curcls)
current_classes = current_classes_int
overlap_ranges = np.zeros([3, 3, len(current_classes)])
for i, curcls in enumerate(current_classes):
overlap_ranges[:, :, i] = np.array(class_to_range[curcls])[:,
np.newaxis]
result = ''
# check whether alpha is valid
compute_aos = False
for anno in dt_annos:
if anno['alpha'].shape[0] != 0:
if anno['alpha'][0] != -10:
compute_aos = True
break
mAPbbox, mAPbev, mAP3d, mAPaos = do_coco_style_eval(
gt_annos, dt_annos, current_classes, overlap_ranges, compute_aos)
for j, curcls in enumerate(current_classes):
# mAP threshold array: [num_minoverlap, metric, class]
# mAP result: [num_class, num_diff, num_minoverlap]
o_range = np.array(class_to_range[curcls])[[0, 2, 1]]
o_range[1] = (o_range[2] - o_range[0]) / (o_range[1] - 1)
result += print_str((f'{class_to_name[curcls]} '
'coco AP@{:.2f}:{:.2f}:{:.2f}:'.format(*o_range)))
result += print_str((f'bbox AP:{mAPbbox[j, 0]:.2f}, '
f'{mAPbbox[j, 1]:.2f}, '
f'{mAPbbox[j, 2]:.2f}'))
result += print_str((f'bev AP:{mAPbev[j, 0]:.2f}, '
f'{mAPbev[j, 1]:.2f}, '
f'{mAPbev[j, 2]:.2f}'))
result += print_str((f'3d AP:{mAP3d[j, 0]:.2f}, '
f'{mAP3d[j, 1]:.2f}, '
f'{mAP3d[j, 2]:.2f}'))
if compute_aos:
result += print_str((f'aos AP:{mAPaos[j, 0]:.2f}, '
f'{mAPaos[j, 1]:.2f}, '
f'{mAPaos[j, 2]:.2f}'))
return result
mmdet3d/core/evaluation/kitti_utils/rotate_iou.py 0 → 100644
View file @ d1aac35d
#####################
# Based on https://github.com/hongzhenwang/RRPN-revise
# Licensed under The MIT License
# Author: yanyan, scrin@foxmail.com
#####################
import math
import numba
import numpy as np
from numba import cuda
@numba.jit(nopython=True)
def div_up(m, n):
return m // n + (m % n > 0)
@cuda.jit('(float32[:], float32[:], float32[:])', device=True, inline=True)
def trangle_area(a, b, c):
return ((a[0] - c[0]) * (b[1] - c[1]) - (a[1] - c[1]) *
(b[0] - c[0])) / 2.0
@cuda.jit('(float32[:], int32)', device=True, inline=True)
def area(int_pts, num_of_inter):
area_val = 0.0
for i in range(num_of_inter - 2):
area_val += abs(
trangle_area(int_pts[:2], int_pts[2 * i + 2:2 * i + 4],
int_pts[2 * i + 4:2 * i + 6]))
return area_val
@cuda.jit('(float32[:], int32)', device=True, inline=True)
def sort_vertex_in_convex_polygon(int_pts, num_of_inter):
if num_of_inter > 0:
center = cuda.local.array((2, ), dtype=numba.float32)
center[:] = 0.0
for i in range(num_of_inter):
center[0] += int_pts[2 * i]
center[1] += int_pts[2 * i + 1]
center[0] /= num_of_inter
center[1] /= num_of_inter
v = cuda.local.array((2, ), dtype=numba.float32)
vs = cuda.local.array((16, ), dtype=numba.float32)
for i in range(num_of_inter):
v[0] = int_pts[2 * i] - center[0]
v[1] = int_pts[2 * i + 1] - center[1]
d = math.sqrt(v[0] * v[0] + v[1] * v[1])
v[0] = v[0] / d
v[1] = v[1] / d
if v[1] < 0:
v[0] = -2 - v[0]
vs[i] = v[0]
j = 0
temp = 0
for i in range(1, num_of_inter):
if vs[i - 1] > vs[i]:
temp = vs[i]
tx = int_pts[2 * i]
ty = int_pts[2 * i + 1]
j = i
while j > 0 and vs[j - 1] > temp:
vs[j] = vs[j - 1]
int_pts[j * 2] = int_pts[j * 2 - 2]
int_pts[j * 2 + 1] = int_pts[j * 2 - 1]
j -= 1
vs[j] = temp
int_pts[j * 2] = tx
int_pts[j * 2 + 1] = ty
@cuda.jit(
'(float32[:], float32[:], int32, int32, float32[:])',
device=True,
inline=True)
def line_segment_intersection(pts1, pts2, i, j, temp_pts):
A = cuda.local.array((2, ), dtype=numba.float32)
B = cuda.local.array((2, ), dtype=numba.float32)
C = cuda.local.array((2, ), dtype=numba.float32)
D = cuda.local.array((2, ), dtype=numba.float32)
A[0] = pts1[2 * i]
A[1] = pts1[2 * i + 1]
B[0] = pts1[2 * ((i + 1) % 4)]
B[1] = pts1[2 * ((i + 1) % 4) + 1]
C[0] = pts2[2 * j]
C[1] = pts2[2 * j + 1]
D[0] = pts2[2 * ((j + 1) % 4)]
D[1] = pts2[2 * ((j + 1) % 4) + 1]
BA0 = B[0] - A[0]
BA1 = B[1] - A[1]
DA0 = D[0] - A[0]
CA0 = C[0] - A[0]
DA1 = D[1] - A[1]
CA1 = C[1] - A[1]
acd = DA1 * CA0 > CA1 * DA0
bcd = (D[1] - B[1]) * (C[0] - B[0]) > (C[1] - B[1]) * (D[0] - B[0])
if acd != bcd:
abc = CA1 * BA0 > BA1 * CA0
abd = DA1 * BA0 > BA1 * DA0
if abc != abd:
DC0 = D[0] - C[0]
DC1 = D[1] - C[1]
ABBA = A[0] * B[1] - B[0] * A[1]
CDDC = C[0] * D[1] - D[0] * C[1]
DH = BA1 * DC0 - BA0 * DC1
Dx = ABBA * DC0 - BA0 * CDDC
Dy = ABBA * DC1 - BA1 * CDDC
temp_pts[0] = Dx / DH
temp_pts[1] = Dy / DH
return True
return False
@cuda.jit(
'(float32[:], float32[:], int32, int32, float32[:])',
device=True,
inline=True)
def line_segment_intersection_v1(pts1, pts2, i, j, temp_pts):
a = cuda.local.array((2, ), dtype=numba.float32)
b = cuda.local.array((2, ), dtype=numba.float32)
c = cuda.local.array((2, ), dtype=numba.float32)
d = cuda.local.array((2, ), dtype=numba.float32)
a[0] = pts1[2 * i]
a[1] = pts1[2 * i + 1]
b[0] = pts1[2 * ((i + 1) % 4)]
b[1] = pts1[2 * ((i + 1) % 4) + 1]
c[0] = pts2[2 * j]
c[1] = pts2[2 * j + 1]
d[0] = pts2[2 * ((j + 1) % 4)]
d[1] = pts2[2 * ((j + 1) % 4) + 1]
area_abc = trangle_area(a, b, c)
area_abd = trangle_area(a, b, d)
if area_abc * area_abd >= 0:
return False
area_cda = trangle_area(c, d, a)
area_cdb = area_cda + area_abc - area_abd
if area_cda * area_cdb >= 0:
return False
t = area_cda / (area_abd - area_abc)
dx = t * (b[0] - a[0])
dy = t * (b[1] - a[1])
temp_pts[0] = a[0] + dx
temp_pts[1] = a[1] + dy
return True
@cuda.jit('(float32, float32, float32[:])', device=True, inline=True)
def point_in_quadrilateral(pt_x, pt_y, corners):
ab0 = corners[2] - corners[0]
ab1 = corners[3] - corners[1]
ad0 = corners[6] - corners[0]
ad1 = corners[7] - corners[1]
ap0 = pt_x - corners[0]
ap1 = pt_y - corners[1]
abab = ab0 * ab0 + ab1 * ab1
abap = ab0 * ap0 + ab1 * ap1
adad = ad0 * ad0 + ad1 * ad1
adap = ad0 * ap0 + ad1 * ap1
return abab >= abap and abap >= 0 and adad >= adap and adap >= 0
@cuda.jit('(float32[:], float32[:], float32[:])', device=True, inline=True)
def quadrilateral_intersection(pts1, pts2, int_pts):
num_of_inter = 0
for i in range(4):
if point_in_quadrilateral(pts1[2 * i], pts1[2 * i + 1], pts2):
int_pts[num_of_inter * 2] = pts1[2 * i]
int_pts[num_of_inter * 2 + 1] = pts1[2 * i + 1]
num_of_inter += 1
if point_in_quadrilateral(pts2[2 * i], pts2[2 * i + 1], pts1):
int_pts[num_of_inter * 2] = pts2[2 * i]
int_pts[num_of_inter * 2 + 1] = pts2[2 * i + 1]
num_of_inter += 1
temp_pts = cuda.local.array((2, ), dtype=numba.float32)
for i in range(4):
for j in range(4):
has_pts = line_segment_intersection(pts1, pts2, i, j, temp_pts)
if has_pts:
int_pts[num_of_inter * 2] = temp_pts[0]
int_pts[num_of_inter * 2 + 1] = temp_pts[1]
num_of_inter += 1
return num_of_inter
@cuda.jit('(float32[:], float32[:])', device=True, inline=True)
def rbbox_to_corners(corners, rbbox):
# generate clockwise corners and rotate it clockwise
angle = rbbox[4]
a_cos = math.cos(angle)
a_sin = math.sin(angle)
center_x = rbbox[0]
center_y = rbbox[1]
x_d = rbbox[2]
y_d = rbbox[3]
corners_x = cuda.local.array((4, ), dtype=numba.float32)
corners_y = cuda.local.array((4, ), dtype=numba.float32)
corners_x[0] = -x_d / 2
corners_x[1] = -x_d / 2
corners_x[2] = x_d / 2
corners_x[3] = x_d / 2
corners_y[0] = -y_d / 2
corners_y[1] = y_d / 2
corners_y[2] = y_d / 2
corners_y[3] = -y_d / 2
for i in range(4):
corners[2 * i] = a_cos * corners_x[i] + a_sin * corners_y[i] + center_x
corners[2 * i +
1] = -a_sin * corners_x[i] + a_cos * corners_y[i] + center_y
@cuda.jit('(float32[:], float32[:])', device=True, inline=True)
def inter(rbbox1, rbbox2):
corners1 = cuda.local.array((8, ), dtype=numba.float32)
corners2 = cuda.local.array((8, ), dtype=numba.float32)
intersection_corners = cuda.local.array((16, ), dtype=numba.float32)
rbbox_to_corners(corners1, rbbox1)
rbbox_to_corners(corners2, rbbox2)
num_intersection = quadrilateral_intersection(corners1, corners2,
intersection_corners)
sort_vertex_in_convex_polygon(intersection_corners, num_intersection)
# print(intersection_corners.reshape([-1, 2])[:num_intersection])
return area(intersection_corners, num_intersection)
@cuda.jit('(float32[:], float32[:], int32)', device=True, inline=True)
def devRotateIoUEval(rbox1, rbox2, criterion=-1):
area1 = rbox1[2] * rbox1[3]
area2 = rbox2[2] * rbox2[3]
area_inter = inter(rbox1, rbox2)
if criterion == -1:
return area_inter / (area1 + area2 - area_inter)
elif criterion == 0:
return area_inter / area1
elif criterion == 1:
return area_inter / area2
else:
return area_inter
@cuda.jit(
'(int64, int64, float32[:], float32[:], float32[:], int32)',
fastmath=False)
def rotate_iou_kernel_eval(N,
K,
dev_boxes,
dev_query_boxes,
dev_iou,
criterion=-1):
threadsPerBlock = 8 * 8
row_start = cuda.blockIdx.x
col_start = cuda.blockIdx.y
tx = cuda.threadIdx.x
row_size = min(N - row_start * threadsPerBlock, threadsPerBlock)
col_size = min(K - col_start * threadsPerBlock, threadsPerBlock)
block_boxes = cuda.shared.array(shape=(64 * 5, ), dtype=numba.float32)
block_qboxes = cuda.shared.array(shape=(64 * 5, ), dtype=numba.float32)
dev_query_box_idx = threadsPerBlock * col_start + tx
dev_box_idx = threadsPerBlock * row_start + tx
if (tx < col_size):
block_qboxes[tx * 5 + 0] = dev_query_boxes[dev_query_box_idx * 5 + 0]
block_qboxes[tx * 5 + 1] = dev_query_boxes[dev_query_box_idx * 5 + 1]
block_qboxes[tx * 5 + 2] = dev_query_boxes[dev_query_box_idx * 5 + 2]
block_qboxes[tx * 5 + 3] = dev_query_boxes[dev_query_box_idx * 5 + 3]
block_qboxes[tx * 5 + 4] = dev_query_boxes[dev_query_box_idx * 5 + 4]
if (tx < row_size):
block_boxes[tx * 5 + 0] = dev_boxes[dev_box_idx * 5 + 0]
block_boxes[tx * 5 + 1] = dev_boxes[dev_box_idx * 5 + 1]
block_boxes[tx * 5 + 2] = dev_boxes[dev_box_idx * 5 + 2]
block_boxes[tx * 5 + 3] = dev_boxes[dev_box_idx * 5 + 3]
block_boxes[tx * 5 + 4] = dev_boxes[dev_box_idx * 5 + 4]
cuda.syncthreads()
if tx < row_size:
for i in range(col_size):
offset = (
row_start * threadsPerBlock * K + col_start * threadsPerBlock +
tx * K + i)
dev_iou[offset] = devRotateIoUEval(block_qboxes[i * 5:i * 5 + 5],
block_boxes[tx * 5:tx * 5 + 5],
criterion)
def rotate_iou_gpu_eval(boxes, query_boxes, criterion=-1, device_id=0):
"""rotated box iou running in gpu. 500x faster than cpu version
(take 5ms in one example with numba.cuda code).
convert from [this project](
https://github.com/hongzhenwang/RRPN-revise/tree/master/lib/rotation).
Args:
boxes (float tensor: [N, 5]): rbboxes. format: centers, dims,
angles(clockwise when positive)
query_boxes (float tensor: [K, 5]): [description]
device_id (int, optional): Defaults to 0. [description]
Returns:
[type]: [description]
"""
boxes = boxes.astype(np.float32)
query_boxes = query_boxes.astype(np.float32)
N = boxes.shape[0]
K = query_boxes.shape[0]
iou = np.zeros((N, K), dtype=np.float32)
if N == 0 or K == 0:
return iou
threadsPerBlock = 8 * 8
cuda.select_device(device_id)
blockspergrid = (div_up(N, threadsPerBlock), div_up(K, threadsPerBlock))
stream = cuda.stream()
with stream.auto_synchronize():
boxes_dev = cuda.to_device(boxes.reshape([-1]), stream)
query_boxes_dev = cuda.to_device(query_boxes.reshape([-1]), stream)
iou_dev = cuda.to_device(iou.reshape([-1]), stream)
rotate_iou_kernel_eval[blockspergrid, threadsPerBlock,
stream](N, K, boxes_dev, query_boxes_dev,
iou_dev, criterion)
iou_dev.copy_to_host(iou.reshape([-1]), stream=stream)
return iou.astype(boxes.dtype)
mmdet3d/core/evaluation/mean_ap.py 0 → 100644
View file @ d1aac35d
import mmcv
import numpy as np
from terminaltables import AsciiTable
from .bbox_overlaps import bbox_overlaps
from .class_names import get_classes
def average_precision(recalls, precisions, mode='area'):
"""Calculate average precision (for single or multiple scales).
Args:
recalls (ndarray): shape (num_scales, num_dets) or (num_dets, )
precisions (ndarray): shape (num_scales, num_dets) or (num_dets, )
mode (str): 'area' or '11points', 'area' means calculating the area
under precision-recall curve, '11points' means calculating
the average precision of recalls at [0, 0.1, ..., 1]
Returns:
float or ndarray: calculated average precision
"""
no_scale = False
if recalls.ndim == 1:
no_scale = True
recalls = recalls[np.newaxis, :]
precisions = precisions[np.newaxis, :]
assert recalls.shape == precisions.shape and recalls.ndim == 2
num_scales = recalls.shape[0]
ap = np.zeros(num_scales, dtype=np.float32)
if mode == 'area':
zeros = np.zeros((num_scales, 1), dtype=recalls.dtype)
ones = np.ones((num_scales, 1), dtype=recalls.dtype)
mrec = np.hstack((zeros, recalls, ones))
mpre = np.hstack((zeros, precisions, zeros))
for i in range(mpre.shape[1] - 1, 0, -1):
mpre[:, i - 1] = np.maximum(mpre[:, i - 1], mpre[:, i])
for i in range(num_scales):
ind = np.where(mrec[i, 1:] != mrec[i, :-1])[0]
ap[i] = np.sum(
(mrec[i, ind + 1] - mrec[i, ind]) * mpre[i, ind + 1])
elif mode == '11points':
for i in range(num_scales):
for thr in np.arange(0, 1 + 1e-3, 0.1):
precs = precisions[i, recalls[i, :] >= thr]
prec = precs.max() if precs.size > 0 else 0
ap[i] += prec
ap /= 11
else:
raise ValueError(
'Unrecognized mode, only "area" and "11points" are supported')
if no_scale:
ap = ap[0]
return ap
def tpfp_imagenet(det_bboxes,
gt_bboxes,
gt_ignore,
default_iou_thr,
area_ranges=None):
"""Check if detected bboxes are true positive or false positive.
Args:
det_bbox (ndarray): the detected bbox
gt_bboxes (ndarray): ground truth bboxes of this image
gt_ignore (ndarray): indicate if gts are ignored for evaluation or not
default_iou_thr (float): the iou thresholds for medium and large bboxes
area_ranges (list or None): gt bbox area ranges
Returns:
tuple: two arrays (tp, fp) whose elements are 0 and 1
"""
num_dets = det_bboxes.shape[0]
num_gts = gt_bboxes.shape[0]
if area_ranges is None:
area_ranges = [(None, None)]
num_scales = len(area_ranges)
# tp and fp are of shape (num_scales, num_gts), each row is tp or fp
# of a certain scale.
tp = np.zeros((num_scales, num_dets), dtype=np.float32)
fp = np.zeros((num_scales, num_dets), dtype=np.float32)
if gt_bboxes.shape[0] == 0:
if area_ranges == [(None, None)]:
fp[...] = 1
else:
det_areas = (det_bboxes[:, 2] - det_bboxes[:, 0]) * (
det_bboxes[:, 3] - det_bboxes[:, 1])
for i, (min_area, max_area) in enumerate(area_ranges):
fp[i, (det_areas >= min_area) & (det_areas < max_area)] = 1
return tp, fp
ious = bbox_overlaps(det_bboxes, gt_bboxes - 1)
gt_w = gt_bboxes[:, 2] - gt_bboxes[:, 0]
gt_h = gt_bboxes[:, 3] - gt_bboxes[:, 1]
iou_thrs = np.minimum((gt_w * gt_h) / ((gt_w + 10.0) * (gt_h + 10.0)),
default_iou_thr)
# sort all detections by scores in descending order
sort_inds = np.argsort(-det_bboxes[:, -1])
for k, (min_area, max_area) in enumerate(area_ranges):
gt_covered = np.zeros(num_gts, dtype=bool)
# if no area range is specified, gt_area_ignore is all False
if min_area is None:
gt_area_ignore = np.zeros_like(gt_ignore, dtype=bool)
else:
gt_areas = gt_w * gt_h
gt_area_ignore = (gt_areas < min_area) | (gt_areas >= max_area)
for i in sort_inds:
max_iou = -1
matched_gt = -1
# find best overlapped available gt
for j in range(num_gts):
# different from PASCAL VOC: allow finding other gts if the
# best overlaped ones are already matched by other det bboxes
if gt_covered[j]:
continue
elif ious[i, j] >= iou_thrs[j] and ious[i, j] > max_iou:
max_iou = ious[i, j]
matched_gt = j
# there are 4 cases for a det bbox:
# 1. it matches a gt, tp = 1, fp = 0
# 2. it matches an ignored gt, tp = 0, fp = 0
# 3. it matches no gt and within area range, tp = 0, fp = 1
# 4. it matches no gt but is beyond area range, tp = 0, fp = 0
if matched_gt >= 0:
gt_covered[matched_gt] = 1
if not (gt_ignore[matched_gt] or gt_area_ignore[matched_gt]):
tp[k, i] = 1
elif min_area is None:
fp[k, i] = 1
else:
bbox = det_bboxes[i, :4]
area = (bbox[2] - bbox[0]) * (bbox[3] - bbox[1])
if area >= min_area and area < max_area:
fp[k, i] = 1
return tp, fp
def tpfp_default(det_bboxes, gt_bboxes, gt_ignore, iou_thr, area_ranges=None):
"""Check if detected bboxes are true positive or false positive.
Args:
det_bbox (ndarray): the detected bbox
gt_bboxes (ndarray): ground truth bboxes of this image
gt_ignore (ndarray): indicate if gts are ignored for evaluation or not
iou_thr (float): the iou thresholds
Returns:
tuple: (tp, fp), two arrays whose elements are 0 and 1
"""
num_dets = det_bboxes.shape[0]
num_gts = gt_bboxes.shape[0]
if area_ranges is None:
area_ranges = [(None, None)]
num_scales = len(area_ranges)
# tp and fp are of shape (num_scales, num_gts), each row is tp or fp of
# a certain scale
tp = np.zeros((num_scales, num_dets), dtype=np.float32)
fp = np.zeros((num_scales, num_dets), dtype=np.float32)
# if there is no gt bboxes in this image, then all det bboxes
# within area range are false positives
if gt_bboxes.shape[0] == 0:
if area_ranges == [(None, None)]:
fp[...] = 1
else:
det_areas = (det_bboxes[:, 2] - det_bboxes[:, 0]) * (
det_bboxes[:, 3] - det_bboxes[:, 1])
for i, (min_area, max_area) in enumerate(area_ranges):
fp[i, (det_areas >= min_area) & (det_areas < max_area)] = 1
return tp, fp
ious = bbox_overlaps(det_bboxes, gt_bboxes)
ious_max = ious.max(axis=1)
ious_argmax = ious.argmax(axis=1)
sort_inds = np.argsort(-det_bboxes[:, -1])
for k, (min_area, max_area) in enumerate(area_ranges):
gt_covered = np.zeros(num_gts, dtype=bool)
# if no area range is specified, gt_area_ignore is all False
if min_area is None:
gt_area_ignore = np.zeros_like(gt_ignore, dtype=bool)
else:
gt_areas = (gt_bboxes[:, 2] - gt_bboxes[:, 0]) * (
gt_bboxes[:, 3] - gt_bboxes[:, 1])
gt_area_ignore = (gt_areas < min_area) | (gt_areas >= max_area)
for i in sort_inds:
if ious_max[i] >= iou_thr:
matched_gt = ious_argmax[i]
if not (gt_ignore[matched_gt] or gt_area_ignore[matched_gt]):
if not gt_covered[matched_gt]:
gt_covered[matched_gt] = True
tp[k, i] = 1
else:
fp[k, i] = 1
# otherwise ignore this detected bbox, tp = 0, fp = 0
elif min_area is None:
fp[k, i] = 1
else:
bbox = det_bboxes[i, :4]
area = (bbox[2] - bbox[0]) * (bbox[3] - bbox[1])
if area >= min_area and area < max_area:
fp[k, i] = 1
return tp, fp
def get_cls_results(det_results, gt_bboxes, gt_labels, gt_ignore, class_id):
"""Get det results and gt information of a certain class."""
cls_dets = [det[class_id]
for det in det_results] # det bboxes of this class
cls_gts = [] # gt bboxes of this class
cls_gt_ignore = []
for j in range(len(gt_bboxes)):
gt_bbox = gt_bboxes[j]
cls_inds = (gt_labels[j] == class_id)
cls_gt = gt_bbox[cls_inds, :] if gt_bbox.shape[0] > 0 else gt_bbox
cls_gts.append(cls_gt)
if gt_ignore is None:
cls_gt_ignore.append(np.zeros(cls_gt.shape[0], dtype=np.int32))
else:
cls_gt_ignore.append(gt_ignore[j][cls_inds])
return cls_dets, cls_gts, cls_gt_ignore
def eval_map(det_results,
gt_bboxes,
gt_labels,
gt_ignore=None,
scale_ranges=None,
iou_thr=0.5,
dataset=None,
print_summary=True):
"""Evaluate mAP of a dataset.
Args:
det_results (list): a list of list, [[cls1_det, cls2_det, ...], ...]
gt_bboxes (list): ground truth bboxes of each image, a list of K*4
array.
gt_labels (list): ground truth labels of each image, a list of K array
gt_ignore (list): gt ignore indicators of each image, a list of K array
scale_ranges (list, optional): [(min1, max1), (min2, max2), ...]
iou_thr (float): IoU threshold
dataset (None or str or list): dataset name or dataset classes, there
are minor differences in metrics for different datsets, e.g.
"voc07", "imagenet_det", etc.
print_summary (bool): whether to print the mAP summary
Returns:
tuple: (mAP, [dict, dict, ...])
"""
assert len(det_results) == len(gt_bboxes) == len(gt_labels)
if gt_ignore is not None:
assert len(gt_ignore) == len(gt_labels)
for i in range(len(gt_ignore)):
assert len(gt_labels[i]) == len(gt_ignore[i])
area_ranges = ([(rg[0]**2, rg[1]**2) for rg in scale_ranges]
if scale_ranges is not None else None)
num_scales = len(scale_ranges) if scale_ranges is not None else 1
eval_results = []
num_classes = len(det_results[0]) # positive class num
gt_labels = [
label if label.ndim == 1 else label[:, 0] for label in gt_labels
]
for i in range(num_classes):
# get gt and det bboxes of this class
cls_dets, cls_gts, cls_gt_ignore = get_cls_results(
det_results, gt_bboxes, gt_labels, gt_ignore, i)
# calculate tp and fp for each image
tpfp_func = (
tpfp_imagenet if dataset in ['det', 'vid'] else tpfp_default)
tpfp = [
tpfp_func(cls_dets[j], cls_gts[j], cls_gt_ignore[j], iou_thr,
area_ranges) for j in range(len(cls_dets))
]
tp, fp = tuple(zip(*tpfp))
# calculate gt number of each scale, gts ignored or beyond scale
# are not counted
num_gts = np.zeros(num_scales, dtype=int)
for j, bbox in enumerate(cls_gts):
if area_ranges is None:
num_gts[0] += np.sum(np.logical_not(cls_gt_ignore[j]))
else:
gt_areas = (bbox[:, 2] - bbox[:, 0]) * (
bbox[:, 3] - bbox[:, 1])
for k, (min_area, max_area) in enumerate(area_ranges):
num_gts[k] += np.sum(
np.logical_not(cls_gt_ignore[j])
& (gt_areas >= min_area) & (gt_areas < max_area))
# sort all det bboxes by score, also sort tp and fp
cls_dets = np.vstack(cls_dets)
num_dets = cls_dets.shape[0]
sort_inds = np.argsort(-cls_dets[:, -1])
tp = np.hstack(tp)[:, sort_inds]
fp = np.hstack(fp)[:, sort_inds]
# calculate recall and precision with tp and fp
tp = np.cumsum(tp, axis=1)
fp = np.cumsum(fp, axis=1)
eps = np.finfo(np.float32).eps
recalls = tp / np.maximum(num_gts[:, np.newaxis], eps)
precisions = tp / np.maximum((tp + fp), eps)
# calculate AP
if scale_ranges is None:
recalls = recalls[0, :]
precisions = precisions[0, :]
num_gts = num_gts.item()
mode = 'area' if dataset != 'voc07' else '11points'
ap = average_precision(recalls, precisions, mode)
eval_results.append({
'num_gts': num_gts,
'num_dets': num_dets,
'recall': recalls,
'precision': precisions,
'ap': ap
})
if scale_ranges is not None:
# shape (num_classes, num_scales)
all_ap = np.vstack([cls_result['ap'] for cls_result in eval_results])
all_num_gts = np.vstack(
[cls_result['num_gts'] for cls_result in eval_results])
mean_ap = []
for i in range(num_scales):
if np.any(all_num_gts[:, i] > 0):
mean_ap.append(all_ap[all_num_gts[:, i] > 0, i].mean())
else:
mean_ap.append(0.0)
else:
aps = []
for cls_result in eval_results:
if cls_result['num_gts'] > 0:
aps.append(cls_result['ap'])
mean_ap = np.array(aps).mean().item() if aps else 0.0
if print_summary:
print_map_summary(mean_ap, eval_results, dataset, area_ranges)
return mean_ap, eval_results
def print_map_summary(mean_ap, results, dataset=None, ranges=None):
"""Print mAP and results of each class.
Args:
mean_ap(float): calculated from `eval_map`
results(list): calculated from `eval_map`
dataset(None or str or list): dataset name or dataset classes.
ranges(list or Tuple): ranges of areas
"""
num_scales = len(results[0]['ap']) if isinstance(results[0]['ap'],
np.ndarray) else 1
if ranges is not None:
assert len(ranges) == num_scales
num_classes = len(results)
recalls = np.zeros((num_scales, num_classes), dtype=np.float32)
precisions = np.zeros((num_scales, num_classes), dtype=np.float32)
aps = np.zeros((num_scales, num_classes), dtype=np.float32)
num_gts = np.zeros((num_scales, num_classes), dtype=int)
for i, cls_result in enumerate(results):
if cls_result['recall'].size > 0:
recalls[:, i] = np.array(cls_result['recall'], ndmin=2)[:, -1]
precisions[:, i] = np.array(
cls_result['precision'], ndmin=2)[:, -1]
aps[:, i] = cls_result['ap']
num_gts[:, i] = cls_result['num_gts']
if dataset is None:
label_names = [str(i) for i in range(0, num_classes)]
elif mmcv.is_str(dataset):
label_names = get_classes(dataset)
else:
label_names = dataset
if not isinstance(mean_ap, list):
mean_ap = [mean_ap]
header = ['class', 'gts', 'dets', 'recall', 'precision', 'ap']
for i in range(num_scales):
if ranges is not None:
print('Area range ', ranges[i])
table_data = [header]
for j in range(num_classes):
row_data = [
label_names[j], num_gts[i, j], results[j]['num_dets'],
'{:.3f}'.format(recalls[i, j]),
'{:.3f}'.format(precisions[i, j]), '{:.3f}'.format(aps[i, j])
]
table_data.append(row_data)
table_data.append(['mAP', '', '', '', '', '{:.3f}'.format(mean_ap[i])])
table = AsciiTable(table_data)
table.inner_footing_row_border = True
print(table.table)
mmdet3d/core/evaluation/recall.py 0 → 100644
View file @ d1aac35d
import numpy as np
from terminaltables import AsciiTable
from ..bbox import bbox_overlaps_2d
def _recalls(all_ious, proposal_nums, thrs):
img_num = all_ious.shape[0]
total_gt_num = sum([ious.shape[0] for ious in all_ious])
_ious = np.zeros((proposal_nums.size, total_gt_num), dtype=np.float32)
for k, proposal_num in enumerate(proposal_nums):
tmp_ious = np.zeros(0)
for i in range(img_num):
ious = all_ious[i][:, :proposal_num].copy()
gt_ious = np.zeros((ious.shape[0]))
if ious.size == 0:
tmp_ious = np.hstack((tmp_ious, gt_ious))
continue
for j in range(ious.shape[0]):
gt_max_overlaps = ious.argmax(axis=1)
max_ious = ious[np.arange(0, ious.shape[0]), gt_max_overlaps]
gt_idx = max_ious.argmax()
gt_ious[j] = max_ious[gt_idx]
box_idx = gt_max_overlaps[gt_idx]
ious[gt_idx, :] = -1
ious[:, box_idx] = -1
tmp_ious = np.hstack((tmp_ious, gt_ious))
_ious[k, :] = tmp_ious
_ious = np.fliplr(np.sort(_ious, axis=1))
recalls = np.zeros((proposal_nums.size, thrs.size))
for i, thr in enumerate(thrs):
recalls[:, i] = (_ious >= thr).sum(axis=1) / float(total_gt_num)
return recalls
def set_recall_param(proposal_nums, iou_thrs):
"""Check proposal_nums and iou_thrs and set correct format.
"""
if isinstance(proposal_nums, list):
_proposal_nums = np.array(proposal_nums)
elif isinstance(proposal_nums, int):
_proposal_nums = np.array([proposal_nums])
else:
_proposal_nums = proposal_nums
if iou_thrs is None:
_iou_thrs = np.array([0.5])
elif isinstance(iou_thrs, list):
_iou_thrs = np.array(iou_thrs)
elif isinstance(iou_thrs, float):
_iou_thrs = np.array([iou_thrs])
else:
_iou_thrs = iou_thrs
return _proposal_nums, _iou_thrs
def eval_recalls(gts,
proposals,
proposal_nums=None,
iou_thrs=None,
print_summary=True):
"""Calculate recalls.
Args:
gts(list or ndarray): a list of arrays of shape (n, 4)
proposals(list or ndarray): a list of arrays of shape (k, 4) or (k, 5)
proposal_nums(int or list of int or ndarray): top N proposals
thrs(float or list or ndarray): iou thresholds
Returns:
ndarray: recalls of different ious and proposal nums
"""
img_num = len(gts)
assert img_num == len(proposals)
proposal_nums, iou_thrs = set_recall_param(proposal_nums, iou_thrs)
all_ious = []
for i in range(img_num):
if proposals[i].ndim == 2 and proposals[i].shape[1] == 5:
scores = proposals[i][:, 4]
sort_idx = np.argsort(scores)[::-1]
img_proposal = proposals[i][sort_idx, :]
else:
img_proposal = proposals[i]
prop_num = min(img_proposal.shape[0], proposal_nums[-1])
if gts[i] is None or gts[i].shape[0] == 0:
ious = np.zeros((0, img_proposal.shape[0]), dtype=np.float32)
else:
ious = bbox_overlaps_2d(gts[i], img_proposal[:prop_num, :4])
all_ious.append(ious)
all_ious = np.array(all_ious)
recalls = _recalls(all_ious, proposal_nums, iou_thrs)
if print_summary:
print_recall_summary(recalls, proposal_nums, iou_thrs)
return recalls
def print_recall_summary(recalls,
proposal_nums,
iou_thrs,
row_idxs=None,
col_idxs=None):
"""Print recalls in a table.
Args:
recalls(ndarray): calculated from `bbox_recalls`
proposal_nums(ndarray or list): top N proposals
iou_thrs(ndarray or list): iou thresholds
row_idxs(ndarray): which rows(proposal nums) to print
col_idxs(ndarray): which cols(iou thresholds) to print
"""
proposal_nums = np.array(proposal_nums, dtype=np.int32)
iou_thrs = np.array(iou_thrs)
if row_idxs is None:
row_idxs = np.arange(proposal_nums.size)
if col_idxs is None:
col_idxs = np.arange(iou_thrs.size)
row_header = [''] + iou_thrs[col_idxs].tolist()
table_data = [row_header]
for i, num in enumerate(proposal_nums[row_idxs]):
row = [
'{:.3f}'.format(val)
for val in recalls[row_idxs[i], col_idxs].tolist()
]
row.insert(0, num)
table_data.append(row)
table = AsciiTable(table_data)
print(table.table)
def plot_num_recall(recalls, proposal_nums):
"""Plot Proposal_num-Recalls curve.
Args:
recalls(ndarray or list): shape (k,)
proposal_nums(ndarray or list): same shape as `recalls`
"""
if isinstance(proposal_nums, np.ndarray):
_proposal_nums = proposal_nums.tolist()
else:
_proposal_nums = proposal_nums
if isinstance(recalls, np.ndarray):
_recalls = recalls.tolist()
else:
_recalls = recalls
import matplotlib.pyplot as plt
f = plt.figure()
plt.plot([0] + _proposal_nums, [0] + _recalls)
plt.xlabel('Proposal num')
plt.ylabel('Recall')
plt.axis([0, proposal_nums.max(), 0, 1])
f.show()
def plot_iou_recall(recalls, iou_thrs):
"""Plot IoU-Recalls curve.
Args:
recalls(ndarray or list): shape (k,)
iou_thrs(ndarray or list): same shape as `recalls`
"""
if isinstance(iou_thrs, np.ndarray):
_iou_thrs = iou_thrs.tolist()
else:
_iou_thrs = iou_thrs
if isinstance(recalls, np.ndarray):
_recalls = recalls.tolist()
else:
_recalls = recalls
import matplotlib.pyplot as plt
f = plt.figure()
plt.plot(_iou_thrs + [1.0], _recalls + [0.])
plt.xlabel('IoU')
plt.ylabel('Recall')
plt.axis([iou_thrs.min(), 1, 0, 1])
f.show()
mmdet3d/core/optimizer/__init__.py 0 → 100644
View file @ d1aac35d
from .builder import build_optimizer
from .mix_optimizer import MixedOptimizer
from .registry import OPTIMIZERS
__all__ = ['OPTIMIZERS', 'build_optimizer', 'MixedOptimizer']
mmdet3d/core/optimizer/builder.py 0 → 100644
View file @ d1aac35d
import re
import torch
from mmdet.utils import build_from_cfg, get_root_logger
from .registry import OPTIMIZERS
def build_optimizer(model, optimizer_cfg):
"""Build optimizer from configs.
Args:
model (:obj:`nn.Module`): The model with parameters to be optimized.
optimizer_cfg (dict): The config dict of the optimizer.
Positional fields are:
- type: class name of the optimizer.
- lr: base learning rate.
Optional fields are:
- any arguments of the corresponding optimizer type, e.g.,
weight_decay, momentum, etc.
- paramwise_options: a dict with 4 accepted fileds
(bias_lr_mult, bias_decay_mult, norm_decay_mult,
dwconv_decay_mult).
`bias_lr_mult` and `bias_decay_mult` will be multiplied to
the lr and weight decay respectively for all bias parameters
(except for the normalization layers), and
`norm_decay_mult` will be multiplied to the weight decay
for all weight and bias parameters of normalization layers.
`dwconv_decay_mult` will be multiplied to the weight decay
for all weight and bias parameters of depthwise conv layers.
Returns:
torch.optim.Optimizer: The initialized optimizer.
Example:
>>> import torch
>>> model = torch.nn.modules.Conv1d(1, 1, 1)
>>> optimizer_cfg = dict(type='SGD', lr=0.01, momentum=0.9,
>>> weight_decay=0.0001)
>>> optimizer = build_optimizer(model, optimizer_cfg)
"""
if hasattr(model, 'module'):
model = model.module
optimizer_cfg = optimizer_cfg.copy()
if isinstance(optimizer_cfg, list):
# Assume paramwise_options is None if optimizer_cfg is list
from .mix_optimizer import MixedOptimizer
logger = get_root_logger()
keys = [optimizer.pop('key') for optimizer in optimizer_cfg]
keys_params = {key: [] for key in keys}
keys_params_name = {key: [] for key in keys}
keys_optimizer = []
for name, param in model.named_parameters():
param_group = {'params': [param]}
find_flag = False
for key in keys:
if key in name:
keys_params[key].append(param_group)
keys_params_name[key].append(name)
find_flag = True
break
assert find_flag, 'key {} is not matched to any optimizer'.format(
name)
step_intervals = []
for key, single_cfg in zip(keys, optimizer_cfg):
optimizer_cls = getattr(torch.optim, single_cfg.pop('type'))
step_intervals.append(single_cfg.pop('step_interval', 1))
single_optim = optimizer_cls(keys_params[key], **single_cfg)
keys_optimizer.append(single_optim)
logger.info('{} optimizes key:\n {}\n'.format(
optimizer_cls.__name__, keys_params_name[key]))
mix_optimizer = MixedOptimizer(keys_optimizer, step_intervals)
return mix_optimizer
else:
paramwise_options = optimizer_cfg.pop('paramwise_options', None)
# if no paramwise option is specified, just use the global setting
if paramwise_options is None:
params = model.parameters()
else:
assert isinstance(paramwise_options, dict)
# get base lr and weight decay
base_lr = optimizer_cfg['lr']
base_wd = optimizer_cfg.get('weight_decay', None)
# weight_decay must be explicitly specified if mult is specified
if ('bias_decay_mult' in paramwise_options
or 'norm_decay_mult' in paramwise_options
or 'dwconv_decay_mult' in paramwise_options):
assert base_wd is not None
# get param-wise options
bias_lr_mult = paramwise_options.get('bias_lr_mult', 1.)
bias_decay_mult = paramwise_options.get('bias_decay_mult', 1.)
norm_decay_mult = paramwise_options.get('norm_decay_mult', 1.)
dwconv_decay_mult = paramwise_options.get('dwconv_decay_mult', 1.)
named_modules = dict(model.named_modules())
# set param-wise lr and weight decay
params = []
for name, param in model.named_parameters():
param_group = {'params': [param]}
if not param.requires_grad:
# FP16 training needs to copy gradient/weight between master
# weight copy and model weight, it is convenient to keep all
# parameters here to align with model.parameters()
params.append(param_group)
continue
# for norm layers, overwrite the weight decay of weight and bias
# TODO: obtain the norm layer prefixes dynamically
if re.search(r'(bn|gn)(\d+)?.(weight|bias)', name):
if base_wd is not None:
param_group['weight_decay'] = base_wd * norm_decay_mult
# for other layers, overwrite both lr and weight decay of bias
elif name.endswith('.bias'):
param_group['lr'] = base_lr * bias_lr_mult
if base_wd is not None:
param_group['weight_decay'] = base_wd * bias_decay_mult
module_name = name.replace('.weight', '').replace('.bias', '')
if module_name in named_modules and base_wd is not None:
module = named_modules[module_name]
# if this Conv2d is depthwise Conv2d
if isinstance(module, torch.nn.Conv2d) and \
module.in_channels == module.groups:
param_group['weight_decay'] = base_wd * dwconv_decay_mult
# otherwise use the global settings
params.append(param_group)
optimizer_cfg['params'] = params
return build_from_cfg(optimizer_cfg, OPTIMIZERS)
mmdet3d/core/optimizer/mix_optimizer.py 0 → 100644
View file @ d1aac35d
from torch.optim import Optimizer
from .registry import OPTIMIZERS
@OPTIMIZERS.register_module
class MixedOptimizer(Optimizer):
"""Mixed Optimizer that contains multiple optimizers
This optimizer applies the cocktail optimzation for multi-modality models.
"""
def __init__(self, optimizers, step_intervals=None):
self.optimizers = optimizers
self.param_groups = []
for optimizer in self.optimizers:
self.param_groups += optimizer.param_groups
if not isinstance(step_intervals, list):
step_intervals = [1] * len(self.optimizers)
self.step_intervals = step_intervals
self.num_step_updated = 0
def __getstate__(self):
return {
'num_step_updated':
self.num_step_updated,
'defaults': [optimizer.defaults for optimizer in self.optimizers],
'state': [optimizer.state for optimizer in self.optimizers],
'param_groups':
[optimizer.param_groups for optimizer in self.optimizers],
}
def __setstate__(self, state):
self.__dict__.update(state)
def __repr__(self):
format_string = self.__class__.__name__ + ' (\n'
for optimizer in self.optimizers:
format_string += '\t' + optimizer.__repr__ + ',\n'
format_string += ')'
return format_string
def state_dict(self):
state_dicts = [optimizer.state_dict() for optimizer in self.optimizers]
return {
'num_step_updated':
self.num_step_updated,
'state': [state_dict['state'] for state_dict in state_dicts],
'param_groups':
[state_dict['param_groups'] for state_dict in state_dicts],
}
def load_state_dict(self, state_dict):
r"""Loads the optimizer state.
Arguments:
state_dict (dict): optimizer state. Should be an object returned
from a call to :meth:`state_dict`.
"""
assert len(state_dict['state']) == len(self.optimizers)
assert len(state_dict['param_groups']) == len(self.optimizers)
for i, (single_state, single_param_groups) in enumerate(
zip(state_dict['state'], state_dict['param_groups'])):
single_state_dict = dict(
state=single_state, param_groups=single_param_groups)
self.optimizers[i].load_state_dict(single_state_dict)
self.param_groups = []
for optimizer in self.optimizers:
self.param_groups += optimizer.param_groups
self.num_step_updated = state_dict['num_step_updated']
def zero_grad(self):
r"""Clears the gradients of all optimized :class:`torch.Tensor` s."""
for optimizer in self.optimizers:
optimizer.zero_grad()
def step(self, closure=None):
r"""Performs a single optimization step (parameter update).
Arguments:
closure (callable): A closure that reevaluates the model and
returns the loss. Optional for most optimizers.
"""
loss = None
if closure is not None:
loss = closure()
self.num_step_updated += 1
for step_interval, optimizer in zip(self.step_intervals,
self.optimizers):
if self.num_step_updated % step_interval == 0:
optimizer.step()
return loss
def add_param_group(self, param_group):
raise NotImplementedError
mmdet3d/core/optimizer/registry.py 0 → 100644
View file @ d1aac35d
import inspect
import torch
from mmdet.utils import Registry
OPTIMIZERS = Registry('optimizer')
def register_torch_optimizers():
torch_optimizers = []
for module_name in dir(torch.optim):
if module_name.startswith('__'):
continue
_optim = getattr(torch.optim, module_name)
if inspect.isclass(_optim) and issubclass(_optim,
torch.optim.Optimizer):
OPTIMIZERS.register_module(_optim)
torch_optimizers.append(module_name)
return torch_optimizers
TORCH_OPTIMIZERS = register_torch_optimizers()
mmdet3d/core/post_processing/__init__.py 0 → 100644
View file @ d1aac35d
from .bbox_nms import multiclass_nms
from .merge_augs import (merge_aug_bboxes, merge_aug_masks,
merge_aug_proposals, merge_aug_scores)
__all__ = [
'multiclass_nms', 'merge_aug_proposals', 'merge_aug_bboxes',
'merge_aug_scores', 'merge_aug_masks'
]
mmdet3d/core/post_processing/bbox_nms.py 0 → 100644
View file @ d1aac35d
import torch
from mmdet.ops.nms import nms_wrapper
def multiclass_nms(multi_bboxes,
multi_scores,
score_thr,
nms_cfg,
max_num=-1,
score_factors=None):
"""NMS for multi-class bboxes.
Args:
multi_bboxes (Tensor): shape (n, #class*4) or (n, 4)
multi_scores (Tensor): shape (n, #class+1), where the last column
contains scores of the background class, but this will be ignored.
score_thr (float): bbox threshold, bboxes with scores lower than it
will not be considered.
nms_thr (float): NMS IoU threshold
max_num (int): if there are more than max_num bboxes after NMS,
only top max_num will be kept.
score_factors (Tensor): The factors multiplied to scores before
applying NMS
Returns:
tuple: (bboxes, labels), tensors of shape (k, 5) and (k, 1). Labels
are 0-based.
"""
# scores has num_classes + 1 (last one is BG)
num_classes = multi_scores.shape[1] - 1
bboxes, labels = [], []
nms_cfg_ = nms_cfg.copy()
nms_type = nms_cfg_.pop('type', 'nms')
nms_op = getattr(nms_wrapper, nms_type)
# the fg class id range: [0, num_classes-1]
for i in range(0, num_classes):
cls_inds = multi_scores[:, i] > score_thr
if not cls_inds.any():
continue
# get bboxes and scores of this class
if multi_bboxes.shape[1] == 4:
_bboxes = multi_bboxes[cls_inds, :]
else:
_bboxes = multi_bboxes[cls_inds, i * 4:(i + 1) * 4]
_scores = multi_scores[cls_inds, i]
if score_factors is not None:
_scores *= score_factors[cls_inds]
cls_dets = torch.cat([_bboxes, _scores[:, None]], dim=1)
cls_dets, _ = nms_op(cls_dets, **nms_cfg_)
cls_labels = multi_bboxes.new_full((cls_dets.shape[0], ),
i,
dtype=torch.long)
bboxes.append(cls_dets)
labels.append(cls_labels)
if bboxes:
bboxes = torch.cat(bboxes)
labels = torch.cat(labels)
if bboxes.shape[0] > max_num:
_, inds = bboxes[:, -1].sort(descending=True)
inds = inds[:max_num]
bboxes = bboxes[inds]
labels = labels[inds]
else:
bboxes = multi_bboxes.new_zeros((0, 5))
labels = multi_bboxes.new_zeros((0, ), dtype=torch.long)
return bboxes, labels
mmdet3d/core/post_processing/merge_augs.py 0 → 100644
View file @ d1aac35d
import numpy as np
import torch
from mmdet3d.ops import nms
from ..bbox import bbox_mapping_back
def merge_aug_proposals(aug_proposals, img_metas, rpn_test_cfg):
"""Merge augmented proposals (multiscale, flip, etc.)
Args:
aug_proposals (list[Tensor]): proposals from different testing
schemes, shape (n, 5). Note that they are not rescaled to the
original image size.
img_metas (list[dict]): list of image info dict where each dict has:
'img_shape', 'scale_factor', 'flip', and my also contain
'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
For details on the values of these keys see
`mmdet/datasets/pipelines/formatting.py:Collect`.
rpn_test_cfg (dict): rpn test config.
Returns:
Tensor: shape (n, 4), proposals corresponding to original image scale.
"""
recovered_proposals = []
for proposals, img_info in zip(aug_proposals, img_metas):
img_shape = img_info['img_shape']
scale_factor = img_info['scale_factor']
flip = img_info['flip']
_proposals = proposals.clone()
_proposals[:, :4] = bbox_mapping_back(_proposals[:, :4], img_shape,
scale_factor, flip)
recovered_proposals.append(_proposals)
aug_proposals = torch.cat(recovered_proposals, dim=0)
merged_proposals, _ = nms(aug_proposals, rpn_test_cfg.nms_thr)
scores = merged_proposals[:, 4]
_, order = scores.sort(0, descending=True)
num = min(rpn_test_cfg.max_num, merged_proposals.shape[0])
order = order[:num]
merged_proposals = merged_proposals[order, :]
return merged_proposals
def merge_aug_bboxes(aug_bboxes, aug_scores, img_metas, rcnn_test_cfg):
"""Merge augmented detection bboxes and scores.
Args:
aug_bboxes (list[Tensor]): shape (n, 4*#class)
aug_scores (list[Tensor] or None): shape (n, #class)
img_shapes (list[Tensor]): shape (3, ).
rcnn_test_cfg (dict): rcnn test config.
Returns:
tuple: (bboxes, scores)
"""
recovered_bboxes = []
for bboxes, img_info in zip(aug_bboxes, img_metas):
img_shape = img_info[0]['img_shape']
scale_factor = img_info[0]['scale_factor']
flip = img_info[0]['flip']
bboxes = bbox_mapping_back(bboxes, img_shape, scale_factor, flip)
recovered_bboxes.append(bboxes)
bboxes = torch.stack(recovered_bboxes).mean(dim=0)
if aug_scores is None:
return bboxes
else:
scores = torch.stack(aug_scores).mean(dim=0)
return bboxes, scores
def merge_aug_scores(aug_scores):
"""Merge augmented bbox scores."""
if isinstance(aug_scores[0], torch.Tensor):
return torch.mean(torch.stack(aug_scores), dim=0)
else:
return np.mean(aug_scores, axis=0)
def merge_aug_masks(aug_masks, img_metas, rcnn_test_cfg, weights=None):
"""Merge augmented mask prediction.
Args:
aug_masks (list[ndarray]): shape (n, #class, h, w)
img_shapes (list[ndarray]): shape (3, ).
rcnn_test_cfg (dict): rcnn test config.
Returns:
tuple: (bboxes, scores)
"""
recovered_masks = [
mask if not img_info[0]['flip'] else mask[..., ::-1]
for mask, img_info in zip(aug_masks, img_metas)
]
if weights is None:
merged_masks = np.mean(recovered_masks, axis=0)
else:
merged_masks = np.average(
np.array(recovered_masks), axis=0, weights=np.array(weights))
return merged_masks
mmdet3d/core/utils/__init__.py 0 → 100644
View file @ d1aac35d
from .dist_utils import DistOptimizerHook, allreduce_grads
from .misc import tensor2imgs # merge_batch, merge_hook_batch
from .misc import multi_apply, unmap
__all__ = [
'allreduce_grads',
'DistOptimizerHook',
'multi_apply',
'tensor2imgs',
'unmap', # 'merge_batch', 'merge_hook_batch'
]
mmdet3d/core/utils/contextmanagers.py 0 → 100644
View file @ d1aac35d
import asyncio
import contextlib
import logging
import os
import time
from typing import List
import torch
logger = logging.getLogger(__name__)
DEBUG_COMPLETED_TIME = bool(os.environ.get('DEBUG_COMPLETED_TIME', False))
@contextlib.asynccontextmanager
async def completed(trace_name='',
name='',
sleep_interval=0.05,
streams: List[torch.cuda.Stream] = None):
"""
Async context manager that waits for work to complete on
given CUDA streams.
"""
if not torch.cuda.is_available():
yield
return
stream_before_context_switch = torch.cuda.current_stream()
if not streams:
streams = [stream_before_context_switch]
else:
streams = [s if s else stream_before_context_switch for s in streams]
end_events = [
torch.cuda.Event(enable_timing=DEBUG_COMPLETED_TIME) for _ in streams
]
if DEBUG_COMPLETED_TIME:
start = torch.cuda.Event(enable_timing=True)
stream_before_context_switch.record_event(start)
cpu_start = time.monotonic()
logger.debug('%s %s starting, streams: %s', trace_name, name, streams)
grad_enabled_before = torch.is_grad_enabled()
try:
yield
finally:
current_stream = torch.cuda.current_stream()
assert current_stream == stream_before_context_switch
if DEBUG_COMPLETED_TIME:
cpu_end = time.monotonic()
for i, stream in enumerate(streams):
event = end_events[i]
stream.record_event(event)
grad_enabled_after = torch.is_grad_enabled()
# observed change of torch.is_grad_enabled() during concurrent run of
# async_test_bboxes code
assert grad_enabled_before == grad_enabled_after, \
'Unexpected is_grad_enabled() value change'
are_done = [e.query() for e in end_events]
logger.debug('%s %s completed: %s streams: %s', trace_name, name,
are_done, streams)
with torch.cuda.stream(stream_before_context_switch):
while not all(are_done):
await asyncio.sleep(sleep_interval)
are_done = [e.query() for e in end_events]
logger.debug('%s %s completed: %s streams: %s', trace_name,
name, are_done, streams)
current_stream = torch.cuda.current_stream()
assert current_stream == stream_before_context_switch
if DEBUG_COMPLETED_TIME:
cpu_time = (cpu_end - cpu_start) * 1000
stream_times_ms = ''
for i, stream in enumerate(streams):
elapsed_time = start.elapsed_time(end_events[i])
stream_times_ms += ' {stream} {elapsed_time:.2f} ms'.format(
stream, elapsed_time)
logger.info('{trace_name} {name} cpu_time {cpu_time:.2f} ms',
trace_name, name, cpu_time, stream_times_ms)
@contextlib.asynccontextmanager
async def concurrent(streamqueue: asyncio.Queue,
trace_name='concurrent',
name='stream'):
"""Run code concurrently in different streams.
:param streamqueue: asyncio.Queue instance.
Queue tasks define the pool of streams used for concurrent execution.
"""
if not torch.cuda.is_available():
yield
return
initial_stream = torch.cuda.current_stream()
with torch.cuda.stream(initial_stream):
stream = await streamqueue.get()
assert isinstance(stream, torch.cuda.Stream)
try:
with torch.cuda.stream(stream):
logger.debug('%s %s is starting, stream: %s', trace_name, name,
stream)
yield
current = torch.cuda.current_stream()
assert current == stream
logger.debug('%s %s has finished, stream: %s', trace_name,
name, stream)
finally:
streamqueue.task_done()
streamqueue.put_nowait(stream)
mmdet3d/core/utils/dist_utils.py 0 → 100644
View file @ d1aac35d
from collections import OrderedDict
import torch.distributed as dist
from mmcv.runner import OptimizerHook
from torch._utils import (_flatten_dense_tensors, _take_tensors,
_unflatten_dense_tensors)
def _allreduce_coalesced(tensors, world_size, bucket_size_mb=-1):
if bucket_size_mb > 0:
bucket_size_bytes = bucket_size_mb * 1024 * 1024
buckets = _take_tensors(tensors, bucket_size_bytes)
else:
buckets = OrderedDict()
for tensor in tensors:
tp = tensor.type()
if tp not in buckets:
buckets[tp] = []
buckets[tp].append(tensor)
buckets = buckets.values()
for bucket in buckets:
flat_tensors = _flatten_dense_tensors(bucket)
dist.all_reduce(flat_tensors)
flat_tensors.div_(world_size)
for tensor, synced in zip(
bucket, _unflatten_dense_tensors(flat_tensors, bucket)):
tensor.copy_(synced)
def allreduce_grads(params, coalesce=True, bucket_size_mb=-1):
grads = [
param.grad.data for param in params
if param.requires_grad and param.grad is not None
]
world_size = dist.get_world_size()
if coalesce:
_allreduce_coalesced(grads, world_size, bucket_size_mb)
else:
for tensor in grads:
dist.all_reduce(tensor.div_(world_size))
class DistOptimizerHook(OptimizerHook):
def __init__(self, grad_clip=None, coalesce=True, bucket_size_mb=-1):
self.grad_clip = grad_clip
self.coalesce = coalesce
self.bucket_size_mb = bucket_size_mb
def after_train_iter(self, runner):
runner.optimizer.zero_grad()
runner.outputs['loss'].backward()
# allreduce_grads(runner.model.parameters(), self.coalesce,
# self.bucket_size_mb)
if self.grad_clip is not None:
self.clip_grads(runner.model.parameters())
runner.optimizer.step()
mmdet3d/core/utils/kitti_utils.py 0 → 100644
View file @ d1aac35d
import numpy as np
import scipy
import torch
from scipy.spatial import Delaunay
def in_hull(p, hull):
"""
:param p: (N, K) test points
:param hull: (M, K) M corners of a box
:return (N) bool
"""
try:
if not isinstance(hull, Delaunay):
hull = Delaunay(hull)
flag = hull.find_simplex(p) >= 0
except scipy.spatial.qhull.QhullError:
print('Warning: not a hull %s' % str(hull))
flag = np.zeros(p.shape[0], dtype=np.bool)
return flag
def enlarge_box3d(boxes3d, extra_width):
"""
:param boxes3d: (N, 7) [x, y, z, w, l, h, ry] in LiDAR coords
"""
if isinstance(boxes3d, np.ndarray):
large_boxes3d = boxes3d.copy()
else:
large_boxes3d = boxes3d.clone()
large_boxes3d[:, 3:6] += extra_width * 2
# bugfixed: here should be minus, not addion in LiDAR, 20190508
large_boxes3d[:, 2] -= extra_width
return large_boxes3d
def rotate_pc_along_z(pc, rot_angle):
"""
params pc: (N, 3+C), (N, 3) is in the LiDAR coordinate
params rot_angle: rad scalar
Output pc: updated pc with XYZ rotated
"""
cosval = np.cos(rot_angle)
sinval = np.sin(rot_angle)
rotmat = np.array([[cosval, -sinval], [sinval, cosval]])
pc[:, 0:2] = np.dot(pc[:, 0:2], rotmat)
return pc
def rotate_pc_along_z_torch(pc, rot_angle):
"""
:param pc: (N, 512, 3 + C) in the LiDAR coordinate
:param rot_angle: (N)
:return:
TODO: merge with rotate_pc_along_y_torch in bbox_transform.py
"""
cosa = torch.cos(rot_angle).view(-1, 1) # (N, 1)
sina = torch.sin(rot_angle).view(-1, 1) # (N, 1)
raw_1 = torch.cat([cosa, -sina], dim=1) # (N, 2)
raw_2 = torch.cat([sina, cosa], dim=1) # (N, 2)
R = torch.cat((raw_1.unsqueeze(dim=1), raw_2.unsqueeze(dim=1)),
dim=1) # (N, 2, 2)
pc_temp = pc[:, :, 0:2] # (N, 512, 2)
pc[:, :, 0:2] = torch.matmul(pc_temp, R) # (N, 512, 2)
return pc
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment