Skip to content

GitLab

Unverified Commit 7f977ea3 authored by Au1ong's avatar Au1ong Committed by GitHub
Browse files

Add pyramid data augmentation which mentioned in SE-SSD (#751) 



* add pyramid augmentation

* add pyramid augmentation

* format the modified file
Co-authored-by: default avatarShaoshuai Shi <shaoshuaics@gmail.com>
parent a991105c
Show whitespace changes
Inline Side-by-side
Showing with 506 additions and 100 deletions
pcdet/datasets/augmentor/augmentor_utils.py
View file @ 7f977ea3
...@@ -2,6 +2,7 @@ import numpy as np ...@@ -2,6 +2,7 @@ import numpy as np
import math import math
import copy import copy
from ...utils import common_utils from ...utils import common_utils
from ...utils import box_utils
def random_flip_along_x(gt_boxes, points): def random_flip_along_x(gt_boxes, points):
...@@ -22,6 +23,7 @@ def random_flip_along_x(gt_boxes, points): ...@@ -22,6 +23,7 @@ def random_flip_along_x(gt_boxes, points):
return gt_boxes, points return gt_boxes, points
def random_flip_along_y(gt_boxes, points): def random_flip_along_y(gt_boxes, points):
""" """
Args: Args:
...@@ -40,6 +42,7 @@ def random_flip_along_y(gt_boxes, points): ...@@ -40,6 +42,7 @@ def random_flip_along_y(gt_boxes, points):
return gt_boxes, points return gt_boxes, points
def global_rotation(gt_boxes, points, rot_range): def global_rotation(gt_boxes, points, rot_range):
""" """
Args: Args:
...@@ -60,6 +63,7 @@ def global_rotation(gt_boxes, points, rot_range): ...@@ -60,6 +63,7 @@ def global_rotation(gt_boxes, points, rot_range):
return gt_boxes, points return gt_boxes, points
def global_scaling(gt_boxes, points, scale_range): def global_scaling(gt_boxes, points, scale_range):
""" """
Args: Args:
...@@ -76,6 +80,7 @@ def global_scaling(gt_boxes, points, scale_range): ...@@ -76,6 +80,7 @@ def global_scaling(gt_boxes, points, scale_range):
return gt_boxes, points return gt_boxes, points
def random_image_flip_horizontal(image, depth_map, gt_boxes, calib): def random_image_flip_horizontal(image, depth_map, gt_boxes, calib):
""" """
Performs random horizontal flip augmentation Performs random horizontal flip augmentation
...@@ -115,6 +120,7 @@ def random_image_flip_horizontal(image, depth_map, gt_boxes, calib): ...@@ -115,6 +120,7 @@ def random_image_flip_horizontal(image, depth_map, gt_boxes, calib):
return aug_image, aug_depth_map, aug_gt_boxes return aug_image, aug_depth_map, aug_gt_boxes
def random_translation_along_x(gt_boxes, points, offset_range): def random_translation_along_x(gt_boxes, points, offset_range):
""" """
Args: Args:
...@@ -133,6 +139,7 @@ def random_translation_along_x(gt_boxes, points, offset_range): ...@@ -133,6 +139,7 @@ def random_translation_along_x(gt_boxes, points, offset_range):
return gt_boxes, points return gt_boxes, points
def random_translation_along_y(gt_boxes, points, offset_range): def random_translation_along_y(gt_boxes, points, offset_range):
""" """
Args: Args:
...@@ -151,6 +158,7 @@ def random_translation_along_y(gt_boxes, points, offset_range): ...@@ -151,6 +158,7 @@ def random_translation_along_y(gt_boxes, points, offset_range):
return gt_boxes, points return gt_boxes, points
def random_translation_along_z(gt_boxes, points, offset_range): def random_translation_along_z(gt_boxes, points, offset_range):
""" """
Args: Args:
...@@ -166,6 +174,7 @@ def random_translation_along_z(gt_boxes, points, offset_range): ...@@ -166,6 +174,7 @@ def random_translation_along_z(gt_boxes, points, offset_range):
return gt_boxes, points return gt_boxes, points
def random_local_translation_along_x(gt_boxes, points, offset_range): def random_local_translation_along_x(gt_boxes, points, offset_range):
""" """
Args: Args:
...@@ -188,6 +197,7 @@ def random_local_translation_along_x(gt_boxes, points, offset_range): ...@@ -188,6 +197,7 @@ def random_local_translation_along_x(gt_boxes, points, offset_range):
return gt_boxes, points return gt_boxes, points
def random_local_translation_along_y(gt_boxes, points, offset_range): def random_local_translation_along_y(gt_boxes, points, offset_range):
""" """
Args: Args:
...@@ -210,6 +220,7 @@ def random_local_translation_along_y(gt_boxes, points, offset_range): ...@@ -210,6 +220,7 @@ def random_local_translation_along_y(gt_boxes, points, offset_range):
return gt_boxes, points return gt_boxes, points
def random_local_translation_along_z(gt_boxes, points, offset_range): def random_local_translation_along_z(gt_boxes, points, offset_range):
""" """
Args: Args:
...@@ -229,6 +240,7 @@ def random_local_translation_along_z(gt_boxes, points, offset_range): ...@@ -229,6 +240,7 @@ def random_local_translation_along_z(gt_boxes, points, offset_range):
return gt_boxes, points return gt_boxes, points
def global_frustum_dropout_top(gt_boxes, points, intensity_range): def global_frustum_dropout_top(gt_boxes, points, intensity_range):
""" """
Args: Args:
...@@ -238,13 +250,14 @@ def global_frustum_dropout_top(gt_boxes, points, intensity_range): ...@@ -238,13 +250,14 @@ def global_frustum_dropout_top(gt_boxes, points, intensity_range):
Returns: Returns:
""" """
intensity = np.random.uniform(intensity_range[0], intensity_range[1]) intensity = np.random.uniform(intensity_range[0], intensity_range[1])
# threshold = max - length * uniform(0 ~ 0.2)
threshold = np.max(points[:, 2]) - intensity * (np.max(points[:, 2]) - np.min(points[:, 2])) threshold = np.max(points[:, 2]) - intensity * (np.max(points[:, 2]) - np.min(points[:, 2]))
points = points[points[:,2] < threshold]
gt_boxes = gt_boxes[gt_boxes[:,2] < threshold]
points = points[points[:, 2] < threshold]
gt_boxes = gt_boxes[gt_boxes[:, 2] < threshold]
return gt_boxes, points return gt_boxes, points
def global_frustum_dropout_bottom(gt_boxes, points, intensity_range): def global_frustum_dropout_bottom(gt_boxes, points, intensity_range):
""" """
Args: Args:
...@@ -256,11 +269,12 @@ def global_frustum_dropout_bottom(gt_boxes, points, intensity_range): ...@@ -256,11 +269,12 @@ def global_frustum_dropout_bottom(gt_boxes, points, intensity_range):
intensity = np.random.uniform(intensity_range[0], intensity_range[1]) intensity = np.random.uniform(intensity_range[0], intensity_range[1])
threshold = np.min(points[:, 2]) + intensity * (np.max(points[:, 2]) - np.min(points[:, 2])) threshold = np.min(points[:, 2]) + intensity * (np.max(points[:, 2]) - np.min(points[:, 2]))
points = points[points[:,2] > threshold] points = points[points[:, 2] > threshold]
gt_boxes = gt_boxes[gt_boxes[:,2] > threshold] gt_boxes = gt_boxes[gt_boxes[:, 2] > threshold]
return gt_boxes, points return gt_boxes, points
def global_frustum_dropout_left(gt_boxes, points, intensity_range): def global_frustum_dropout_left(gt_boxes, points, intensity_range):
""" """
Args: Args:
...@@ -272,11 +286,12 @@ def global_frustum_dropout_left(gt_boxes, points, intensity_range): ...@@ -272,11 +286,12 @@ def global_frustum_dropout_left(gt_boxes, points, intensity_range):
intensity = np.random.uniform(intensity_range[0], intensity_range[1]) intensity = np.random.uniform(intensity_range[0], intensity_range[1])
threshold = np.max(points[:, 1]) - intensity * (np.max(points[:, 1]) - np.min(points[:, 1])) threshold = np.max(points[:, 1]) - intensity * (np.max(points[:, 1]) - np.min(points[:, 1]))
points = points[points[:,1] < threshold] points = points[points[:, 1] < threshold]
gt_boxes = gt_boxes[gt_boxes[:,1] < threshold] gt_boxes = gt_boxes[gt_boxes[:, 1] < threshold]
return gt_boxes, points return gt_boxes, points
def global_frustum_dropout_right(gt_boxes, points, intensity_range): def global_frustum_dropout_right(gt_boxes, points, intensity_range):
""" """
Args: Args:
...@@ -288,11 +303,12 @@ def global_frustum_dropout_right(gt_boxes, points, intensity_range): ...@@ -288,11 +303,12 @@ def global_frustum_dropout_right(gt_boxes, points, intensity_range):
intensity = np.random.uniform(intensity_range[0], intensity_range[1]) intensity = np.random.uniform(intensity_range[0], intensity_range[1])
threshold = np.min(points[:, 1]) + intensity * (np.max(points[:, 1]) - np.min(points[:, 1])) threshold = np.min(points[:, 1]) + intensity * (np.max(points[:, 1]) - np.min(points[:, 1]))
points = points[points[:,1] > threshold] points = points[points[:, 1] > threshold]
gt_boxes = gt_boxes[gt_boxes[:,1] > threshold] gt_boxes = gt_boxes[gt_boxes[:, 1] > threshold]
return gt_boxes, points return gt_boxes, points
def local_scaling(gt_boxes, points, scale_range): def local_scaling(gt_boxes, points, scale_range):
""" """
Args: Args:
...@@ -374,6 +390,7 @@ def local_rotation(gt_boxes, points, rot_range): ...@@ -374,6 +390,7 @@ def local_rotation(gt_boxes, points, rot_range):
return gt_boxes, points return gt_boxes, points
def local_frustum_dropout_top(gt_boxes, points, intensity_range): def local_frustum_dropout_top(gt_boxes, points, intensity_range):
""" """
Args: Args:
...@@ -387,12 +404,13 @@ def local_frustum_dropout_top(gt_boxes, points, intensity_range): ...@@ -387,12 +404,13 @@ def local_frustum_dropout_top(gt_boxes, points, intensity_range):
intensity = np.random.uniform(intensity_range[0], intensity_range[1]) intensity = np.random.uniform(intensity_range[0], intensity_range[1])
points_in_box, mask = get_points_in_box(points, box) points_in_box, mask = get_points_in_box(points, box)
threshold = (z + dz/2) - intensity * dz threshold = (z + dz / 2) - intensity * dz
points = points[np.logical_not(np.logical_and(mask, points[:,2] >= threshold))] points = points[np.logical_not(np.logical_and(mask, points[:, 2] >= threshold))]
return gt_boxes, points return gt_boxes, points
def local_frustum_dropout_bottom(gt_boxes, points, intensity_range): def local_frustum_dropout_bottom(gt_boxes, points, intensity_range):
""" """
Args: Args:
...@@ -406,12 +424,13 @@ def local_frustum_dropout_bottom(gt_boxes, points, intensity_range): ...@@ -406,12 +424,13 @@ def local_frustum_dropout_bottom(gt_boxes, points, intensity_range):
intensity = np.random.uniform(intensity_range[0], intensity_range[1]) intensity = np.random.uniform(intensity_range[0], intensity_range[1])
points_in_box, mask = get_points_in_box(points, box) points_in_box, mask = get_points_in_box(points, box)
threshold = (z - dz/2) + intensity * dz threshold = (z - dz / 2) + intensity * dz
points = points[np.logical_not(np.logical_and(mask, points[:,2] <= threshold))] points = points[np.logical_not(np.logical_and(mask, points[:, 2] <= threshold))]
return gt_boxes, points return gt_boxes, points
def local_frustum_dropout_left(gt_boxes, points, intensity_range): def local_frustum_dropout_left(gt_boxes, points, intensity_range):
""" """
Args: Args:
...@@ -425,12 +444,13 @@ def local_frustum_dropout_left(gt_boxes, points, intensity_range): ...@@ -425,12 +444,13 @@ def local_frustum_dropout_left(gt_boxes, points, intensity_range):
intensity = np.random.uniform(intensity_range[0], intensity_range[1]) intensity = np.random.uniform(intensity_range[0], intensity_range[1])
points_in_box, mask = get_points_in_box(points, box) points_in_box, mask = get_points_in_box(points, box)
threshold = (y + dy/2) - intensity * dy threshold = (y + dy / 2) - intensity * dy
points = points[np.logical_not(np.logical_and(mask, points[:,1] >= threshold))] points = points[np.logical_not(np.logical_and(mask, points[:, 1] >= threshold))]
return gt_boxes, points return gt_boxes, points
def local_frustum_dropout_right(gt_boxes, points, intensity_range): def local_frustum_dropout_right(gt_boxes, points, intensity_range):
""" """
Args: Args:
...@@ -444,14 +464,15 @@ def local_frustum_dropout_right(gt_boxes, points, intensity_range): ...@@ -444,14 +464,15 @@ def local_frustum_dropout_right(gt_boxes, points, intensity_range):
intensity = np.random.uniform(intensity_range[0], intensity_range[1]) intensity = np.random.uniform(intensity_range[0], intensity_range[1])
points_in_box, mask = get_points_in_box(points, box) points_in_box, mask = get_points_in_box(points, box)
threshold = (y - dy/2) + intensity * dy threshold = (y - dy / 2) + intensity * dy
points = points[np.logical_not(np.logical_and(mask, points[:,1] <= threshold))] points = points[np.logical_not(np.logical_and(mask, points[:, 1] <= threshold))]
return gt_boxes, points return gt_boxes, points
def get_points_in_box(points, gt_box): def get_points_in_box(points, gt_box):
x, y, z = points[:,0], points[:,1], points[:,2] x, y, z = points[:, 0], points[:, 1], points[:, 2]
cx, cy, cz = gt_box[0], gt_box[1], gt_box[2] cx, cy, cz = gt_box[0], gt_box[1], gt_box[2]
dx, dy, dz, rz = gt_box[3], gt_box[4], gt_box[5], gt_box[6] dx, dy, dz, rz = gt_box[3], gt_box[4], gt_box[5], gt_box[6]
shift_x, shift_y, shift_z = x - cx, y - cy, z - cz shift_x, shift_y, shift_z = x - cx, y - cy, z - cz
...@@ -461,10 +482,202 @@ def get_points_in_box(points, gt_box): ...@@ -461,10 +482,202 @@ def get_points_in_box(points, gt_box):
local_x = shift_x * cosa + shift_y * (-sina) local_x = shift_x * cosa + shift_y * (-sina)
local_y = shift_x * sina + shift_y * cosa local_y = shift_x * sina + shift_y * cosa
mask = np.logical_and(abs(shift_z) <= dz / 2.0, \ mask = np.logical_and(abs(shift_z) <= dz / 2.0,
np.logical_and(abs(local_x) <= dx / 2.0 + MARGIN, \ np.logical_and(abs(local_x) <= dx / 2.0 + MARGIN,
abs(local_y) <= dy / 2.0 + MARGIN )) abs(local_y) <= dy / 2.0 + MARGIN))
points = points[mask] points = points[mask]
return points, mask return points, mask
def get_pyramids(boxes):
pyramid_orders = np.array([
[0, 1, 5, 4],
[4, 5, 6, 7],
[7, 6, 2, 3],
[3, 2, 1, 0],
[1, 2, 6, 5],
[0, 4, 7, 3]
])
boxes_corners = box_utils.boxes_to_corners_3d(boxes).reshape(-1, 24)
pyramid_list = []
for order in pyramid_orders:
# frustum polygon: 5 corners, 5 surfaces
pyramid = np.concatenate((
boxes[:, 0:3],
boxes_corners[:, 3 * order[0]: 3 * order[0] + 3],
boxes_corners[:, 3 * order[1]: 3 * order[1] + 3],
boxes_corners[:, 3 * order[2]: 3 * order[2] + 3],
boxes_corners[:, 3 * order[3]: 3 * order[3] + 3]), axis=1)
pyramid_list.append(pyramid[:, None, :])
pyramids = np.concatenate(pyramid_list, axis=1) # [N, 6, 15], 15=5*3
return pyramids
def one_hot(x, num_class=1):
if num_class is None:
num_class = 1
ohx = np.zeros((len(x), num_class))
ohx[range(len(x)), x] = 1
return ohx
def points_in_pyramids_mask(points, pyramids):
pyramids = pyramids.reshape(-1, 5, 3)
flags = np.zeros((points.shape[0], pyramids.shape[0]), dtype=np.bool)
for i, pyramid in enumerate(pyramids):
flags[:, i] = np.logical_or(flags[:, i], box_utils.in_hull(points[:, 0:3], pyramid))
return flags
def local_pyramid_dropout(gt_boxes, points, dropout_prob, pyramids=None):
if pyramids is None:
pyramids = get_pyramids(gt_boxes).reshape([-1, 6, 5, 3]) # each six surface of boxes: [num_boxes, 6, 15=3*5]
drop_pyramid_indices = np.random.randint(0, 6, (pyramids.shape[0]))
drop_pyramid_one_hot = one_hot(drop_pyramid_indices, num_class=6)
drop_box_mask = np.random.uniform(0, 1, (pyramids.shape[0])) <= dropout_prob
if np.sum(drop_box_mask) != 0:
drop_pyramid_mask = (np.tile(drop_box_mask[:, None], [1, 6]) * drop_pyramid_one_hot) > 0
drop_pyramids = pyramids[drop_pyramid_mask]
point_masks = points_in_pyramids_mask(points, drop_pyramids)
points = points[np.logical_not(point_masks.any(-1))]
# print(drop_box_mask)
pyramids = pyramids[np.logical_not(drop_box_mask)]
return gt_boxes, points, pyramids
def local_pyramid_sparsify(gt_boxes, points, prob, max_num_pts, pyramids=None):
if pyramids is None:
pyramids = get_pyramids(gt_boxes).reshape([-1, 6, 5, 3]) # each six surface of boxes: [num_boxes, 6, 15=3*5]
if pyramids.shape[0] > 0:
sparsity_prob, sparsity_num = prob, max_num_pts
sparsify_pyramid_indices = np.random.randint(0, 6, (pyramids.shape[0]))
sparsify_pyramid_one_hot = one_hot(sparsify_pyramid_indices, num_class=6)
sparsify_box_mask = np.random.uniform(0, 1, (pyramids.shape[0])) <= sparsity_prob
sparsify_pyramid_mask = (np.tile(sparsify_box_mask[:, None], [1, 6]) * sparsify_pyramid_one_hot) > 0
# print(sparsify_box_mask)
pyramid_sampled = pyramids[sparsify_pyramid_mask] # (-1,6,5,3)[(num_sample,6)]
# print(pyramid_sampled.shape)
pyramid_sampled_point_masks = points_in_pyramids_mask(points, pyramid_sampled)
pyramid_sampled_points_num = pyramid_sampled_point_masks.sum(0) # the number of points in each surface pyramid
valid_pyramid_sampled_mask = pyramid_sampled_points_num > sparsity_num # only much than sparsity_num should be sparse
sparsify_pyramids = pyramid_sampled[valid_pyramid_sampled_mask]
if sparsify_pyramids.shape[0] > 0:
point_masks = pyramid_sampled_point_masks[:, valid_pyramid_sampled_mask]
remain_points = points[
np.logical_not(point_masks.any(-1))] # points which outside the down sampling pyramid
to_sparsify_points = [points[point_masks[:, i]] for i in range(point_masks.shape[1])]
sparsified_points = []
for sample in to_sparsify_points:
sampled_indices = np.random.choice(sample.shape[0], size=sparsity_num, replace=False)
sparsified_points.append(sample[sampled_indices])
sparsified_points = np.concatenate(sparsified_points, axis=0)
points = np.concatenate([remain_points, sparsified_points], axis=0)
pyramids = pyramids[np.logical_not(sparsify_box_mask)]
return gt_boxes, points, pyramids
def local_pyramid_swap(gt_boxes, points, prob, max_num_pts, pyramids=None):
def get_points_ratio(points, pyramid):
surface_center = (pyramid[3:6] + pyramid[6:9] + pyramid[9:12] + pyramid[12:]) / 4.0
vector_0, vector_1, vector_2 = pyramid[6:9] - pyramid[3:6], pyramid[12:] - pyramid[3:6], pyramid[0:3] - surface_center
alphas = ((points[:, 0:3] - pyramid[3:6]) * vector_0).sum(-1) / np.power(vector_0, 2).sum()
betas = ((points[:, 0:3] - pyramid[3:6]) * vector_1).sum(-1) / np.power(vector_1, 2).sum()
gammas = ((points[:, 0:3] - surface_center) * vector_2).sum(-1) / np.power(vector_2, 2).sum()
return [alphas, betas, gammas]
def recover_points_by_ratio(points_ratio, pyramid):
alphas, betas, gammas = points_ratio
surface_center = (pyramid[3:6] + pyramid[6:9] + pyramid[9:12] + pyramid[12:]) / 4.0
vector_0, vector_1, vector_2 = pyramid[6:9] - pyramid[3:6], pyramid[12:] - pyramid[3:6], pyramid[0:3] - surface_center
points = (alphas[:, None] * vector_0 + betas[:, None] * vector_1) + pyramid[3:6] + gammas[:, None] * vector_2
return points
def recover_points_intensity_by_ratio(points_intensity_ratio, max_intensity, min_intensity):
return points_intensity_ratio * (max_intensity - min_intensity) + min_intensity
# swap partition
if pyramids is None:
pyramids = get_pyramids(gt_boxes).reshape([-1, 6, 5, 3]) # each six surface of boxes: [num_boxes, 6, 15=3*5]
swap_prob, num_thres = prob, max_num_pts
swap_pyramid_mask = np.random.uniform(0, 1, (pyramids.shape[0])) <= swap_prob
if swap_pyramid_mask.sum() > 0:
point_masks = points_in_pyramids_mask(points, pyramids)
point_nums = point_masks.sum(0).reshape(pyramids.shape[0], -1) # [N, 6]
non_zero_pyramids_mask = point_nums > num_thres # ingore dropout pyramids or highly occluded pyramids
selected_pyramids = non_zero_pyramids_mask * swap_pyramid_mask[:,
None] # selected boxes and all their valid pyramids
# print(selected_pyramids)
if selected_pyramids.sum() > 0:
# get to_swap pyramids
index_i, index_j = np.nonzero(selected_pyramids)
selected_pyramid_indices = [np.random.choice(index_j[index_i == i]) \
if e and (index_i == i).any() else 0 for i, e in
enumerate(swap_pyramid_mask)]
selected_pyramids_mask = selected_pyramids * one_hot(selected_pyramid_indices, num_class=6) == 1
to_swap_pyramids = pyramids[selected_pyramids_mask]
# get swapped pyramids
index_i, index_j = np.nonzero(selected_pyramids_mask)
non_zero_pyramids_mask[selected_pyramids_mask] = False
swapped_index_i = np.array([np.random.choice(np.where(non_zero_pyramids_mask[:, j])[0]) if \
np.where(non_zero_pyramids_mask[:, j])[0].shape[0] > 0 else
index_i[i] for i, j in enumerate(index_j.tolist())])
swapped_indicies = np.concatenate([swapped_index_i[:, None], index_j[:, None]], axis=1)
swapped_pyramids = pyramids[
swapped_indicies[:, 0].astype(np.int32), swapped_indicies[:, 1].astype(np.int32)]
# concat to_swap&swapped pyramids
swap_pyramids = np.concatenate([to_swap_pyramids, swapped_pyramids], axis=0)
swap_point_masks = points_in_pyramids_mask(points, swap_pyramids)
remain_points = points[np.logical_not(swap_point_masks.any(-1))]
# swap pyramids
points_res = []
num_swapped_pyramids = swapped_pyramids.shape[0]
for i in range(num_swapped_pyramids):
to_swap_pyramid = to_swap_pyramids[i]
swapped_pyramid = swapped_pyramids[i]
to_swap_points = points[swap_point_masks[:, i]]
swapped_points = points[swap_point_masks[:, i + num_swapped_pyramids]]
# for intensity transform
to_swap_points_intensity_ratio = (to_swap_points[:, -1:] - to_swap_points[:, -1:].min()) / \
np.clip(
(to_swap_points[:, -1:].max() - to_swap_points[:, -1:].min()),
1e-6, 1)
swapped_points_intensity_ratio = (swapped_points[:, -1:] - swapped_points[:, -1:].min()) / \
np.clip(
(swapped_points[:, -1:].max() - swapped_points[:, -1:].min()),
1e-6, 1)
to_swap_points_ratio = get_points_ratio(to_swap_points, to_swap_pyramid.reshape(15))
swapped_points_ratio = get_points_ratio(swapped_points, swapped_pyramid.reshape(15))
new_to_swap_points = recover_points_by_ratio(swapped_points_ratio, to_swap_pyramid.reshape(15))
new_swapped_points = recover_points_by_ratio(to_swap_points_ratio, swapped_pyramid.reshape(15))
# for intensity transform
new_to_swap_points_intensity = recover_points_intensity_by_ratio(
swapped_points_intensity_ratio, to_swap_points[:, -1:].max(),
to_swap_points[:, -1:].min())
new_swapped_points_intensity = recover_points_intensity_by_ratio(
to_swap_points_intensity_ratio, swapped_points[:, -1:].max(),
swapped_points[:, -1:].min())
# new_to_swap_points = np.concatenate([new_to_swap_points, swapped_points[:, -1:]], axis=1)
# new_swapped_points = np.concatenate([new_swapped_points, to_swap_points[:, -1:]], axis=1)
new_to_swap_points = np.concatenate([new_to_swap_points, new_to_swap_points_intensity], axis=1)
new_swapped_points = np.concatenate([new_swapped_points, new_swapped_points_intensity], axis=1)
points_res.append(new_to_swap_points)
points_res.append(new_swapped_points)
points_res = np.concatenate(points_res, axis=0)
points = np.concatenate([remain_points, points_res], axis=0)
return gt_boxes, points
pcdet/datasets/augmentor/data_augmentor.py
View file @ 7f977ea3
...@@ -203,6 +203,29 @@ class DataAugmentor(object): ...@@ -203,6 +203,29 @@ class DataAugmentor(object):
data_dict['points'] = points data_dict['points'] = points
return data_dict return data_dict
def random_local_pyramid_aug(self, data_dict=None, config=None):
"""
Refer to the paper:
SE-SSD: Self-Ensembling Single-Stage Object Detector From Point Cloud
"""
if data_dict is None:
return partial(self.random_local_pyramid_aug, config=config)
gt_boxes, points = data_dict['gt_boxes'], data_dict['points']
gt_boxes, points, pyramids = augmentor_utils.local_pyramid_dropout(gt_boxes, points, config['DROP_PROB'])
gt_boxes, points, pyramids = augmentor_utils.local_pyramid_sparsify(gt_boxes, points,
config['SPARSIFY_PROB'],
config['SPARSIFY_MAX_NUM'],
pyramids)
gt_boxes, points = augmentor_utils.local_pyramid_swap(gt_boxes, points,
config['SWAP_PROB'],
config['SWAP_MAX_NUM'],
pyramids)
data_dict['gt_boxes'] = gt_boxes
data_dict['points'] = points
return data_dict
def forward(self, data_dict): def forward(self, data_dict):
""" """
Args: Args:
......
tools/cfgs/kitti_models/pointpillar_pyramid_aug.yaml 0 → 100644
View file @ 7f977ea3
CLASS_NAMES: ['Car', 'Pedestrian', 'Cyclist']
DATA_CONFIG:
_BASE_CONFIG_: cfgs/dataset_configs/kitti_dataset.yaml
POINT_CLOUD_RANGE: [0, -39.68, -3, 69.12, 39.68, 1]
DATA_PROCESSOR:
- NAME: mask_points_and_boxes_outside_range
REMOVE_OUTSIDE_BOXES: True
- NAME: shuffle_points
SHUFFLE_ENABLED: {
'train': True,
'test': False
}
- NAME: transform_points_to_voxels
VOXEL_SIZE: [0.16, 0.16, 4]
MAX_POINTS_PER_VOXEL: 32
MAX_NUMBER_OF_VOXELS: {
'train': 16000,
'test': 40000
}
DATA_AUGMENTOR:
DISABLE_AUG_LIST: ['placeholder']
AUG_CONFIG_LIST:
- NAME: gt_sampling
USE_ROAD_PLANE: True
DB_INFO_PATH:
- kitti_dbinfos_train.pkl
PREPARE: {
filter_by_min_points: ['Car:5', 'Pedestrian:5', 'Cyclist:5'],
filter_by_difficulty: [-1],
}
SAMPLE_GROUPS: ['Car:15','Pedestrian:15', 'Cyclist:15']
NUM_POINT_FEATURES: 4
DATABASE_WITH_FAKELIDAR: False
REMOVE_EXTRA_WIDTH: [0.0, 0.0, 0.0]
LIMIT_WHOLE_SCENE: False
- NAME: random_world_flip
ALONG_AXIS_LIST: ['x']
- NAME: random_world_rotation
WORLD_ROT_ANGLE: [-0.78539816, 0.78539816]
- NAME: random_world_scaling
WORLD_SCALE_RANGE: [0.95, 1.05]
- NAME: random_local_pyramid_aug
DROP_PROB: 0.25
SPARSIFY_PROB: 0.05
SPARSIFY_MAX_NUM: 50
SWAP_PROB: 0.1
SWAP_MAX_NUM: 50
MODEL:
NAME: PointPillar
VFE:
NAME: PillarVFE
WITH_DISTANCE: False
USE_ABSLOTE_XYZ: True
USE_NORM: True
NUM_FILTERS: [64]
MAP_TO_BEV:
NAME: PointPillarScatter
NUM_BEV_FEATURES: 64
BACKBONE_2D:
NAME: BaseBEVBackbone
LAYER_NUMS: [3, 5, 5]
LAYER_STRIDES: [2, 2, 2]
NUM_FILTERS: [64, 128, 256]
UPSAMPLE_STRIDES: [1, 2, 4]
NUM_UPSAMPLE_FILTERS: [128, 128, 128]
DENSE_HEAD:
NAME: AnchorHeadSingle
CLASS_AGNOSTIC: False
USE_DIRECTION_CLASSIFIER: True
DIR_OFFSET: 0.78539
DIR_LIMIT_OFFSET: 0.0
NUM_DIR_BINS: 2
ANCHOR_GENERATOR_CONFIG: [
{
'class_name': 'Car',
'anchor_sizes': [[3.9, 1.6, 1.56]],
'anchor_rotations': [0, 1.57],
'anchor_bottom_heights': [-1.78],
'align_center': False,
'feature_map_stride': 2,
'matched_threshold': 0.6,
'unmatched_threshold': 0.45
},
{
'class_name': 'Pedestrian',
'anchor_sizes': [[0.8, 0.6, 1.73]],
'anchor_rotations': [0, 1.57],
'anchor_bottom_heights': [-0.6],
'align_center': False,
'feature_map_stride': 2,
'matched_threshold': 0.5,
'unmatched_threshold': 0.35
},
{
'class_name': 'Cyclist',
'anchor_sizes': [[1.76, 0.6, 1.73]],
'anchor_rotations': [0, 1.57],
'anchor_bottom_heights': [-0.6],
'align_center': False,
'feature_map_stride': 2,
'matched_threshold': 0.5,
'unmatched_threshold': 0.35
}
]
TARGET_ASSIGNER_CONFIG:
NAME: AxisAlignedTargetAssigner
POS_FRACTION: -1.0
SAMPLE_SIZE: 512
NORM_BY_NUM_EXAMPLES: False
MATCH_HEIGHT: False
BOX_CODER: ResidualCoder
LOSS_CONFIG:
LOSS_WEIGHTS: {
'cls_weight': 1.0,
'loc_weight': 2.0,
'dir_weight': 0.2,
'code_weights': [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
}
POST_PROCESSING:
RECALL_THRESH_LIST: [0.3, 0.5, 0.7]
SCORE_THRESH: 0.1
OUTPUT_RAW_SCORE: False
EVAL_METRIC: kitti
NMS_CONFIG:
MULTI_CLASSES_NMS: False
NMS_TYPE: nms_gpu
NMS_THRESH: 0.01
NMS_PRE_MAXSIZE: 4096
NMS_POST_MAXSIZE: 500
OPTIMIZATION:
BATCH_SIZE_PER_GPU: 4
NUM_EPOCHS: 80
OPTIMIZER: adam_onecycle
LR: 0.003
WEIGHT_DECAY: 0.01
MOMENTUM: 0.9
MOMS: [0.95, 0.85]
PCT_START: 0.4
DIV_FACTOR: 10
DECAY_STEP_LIST: [35, 45]
LR_DECAY: 0.1
LR_CLIP: 0.0000001
LR_WARMUP: False
WARMUP_EPOCH: 1
GRAD_NORM_CLIP: 10
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