Modify README and add gen_data.py

README.md

@@ -4,10 +4,11 @@
 
 ## 模型结构
 CenterFace是一种人脸检测算法，采用了轻量级网络mobileNetV2作为主干网络，结合特征金字塔网络（FPN）实现anchor free的人脸检测。
-![Architecture of the CenterFace](Architecture of the CenterFace.png)
+
 
 ## 算法原理
 CenterFace模型是一种基于单阶段人脸检测算法，作者借鉴了CenterNet的思想，将人脸检测转换为标准点问题，根据人脸中心点来回归人脸框的大小和五个标志点。
+![Architecture of the CenterFace](Architecture of the CenterFace.png)
 
 ## 环境配置
 ### Docker（方法一）
@@ -51,7 +52,7 @@ pip3 install -r requirements.txt
 
 ## 数据集
 
-[WIDER_FACE](http://shuoyang1213.me/WIDERFACE/index.html)
+WIDER_FACE：http://shuoyang1213.me/WIDERFACE/index.html
 
 ![datasets](datasets.png)
 
@@ -63,25 +64,76 @@ pip3 install -r requirements.txt
 
 [WIDER Face Testing Images(Tencent Drive)](https://share.weiyun.com/5vSUomP)
 
-annotation文件使用的是coco的格式，可以通过百度网盘下载
-
-[Baidu](https://pan.baidu.com/s/1j_2wggZ3bvCuOAfZvjWqTg) 提取码：f9hh
 
 数据集全部解压后的目录结构如下：
 
 ```
-├── WIDER_train
-│   ├── images
-├── WIDER_test
-│   ├── images
-├── WIDER_val
-│   ├── images
+├── wider_face:  存放数据集根目录
+│   ├── WIDER_train: 训练集解压后的文件目录
+│       └── images:
+│           ├──  0--Parade:         对应该类别的所有图片
+│           ├──  ........
+│           └──  61--Street_Battle: 对应该类别的所有图片
+│   ├── WIDER_val: 验证集解压后的文件目录
+│       └──  images:
+│           ├──  0--Parade:         对应该类别的所有图片
+│           ├──  ........
+│           └──  61--Street_Battle: 对应该类别的所有图片
+│   ├── WIDER_test: 训练集解压后的文件目录
+│       └──  images:
+│           ├──  0--Parade:         对应该类别的所有图片
+│           ├──  ........
+│           └──  61--Street_Battle: 对应该类别的所有图片
+```
+
+解压完成后执行以下步骤：
+1. 将训练图片放置于 ./datasets/images/train的目录下，验证数据放置于./datasets/images/val目录下，存放目录结如下
+```
+├── images
+│   ├── train
+│       ├── 0--Parade
+│       ├──  ........
+│       └──  61--Street_Battle
+│   ├── val
+│       ├── 0--Parade
+│       ├──  ........
+│       └──  61--Street_Battle
+```
+
+2. 如果是使用WIDER_train数据, 可直接将./datasets/labels/train下的train_wider_face.json重命名为train_face.json即可，无需进行标注文件格式转换；反之，需要将训练图片对应的人脸关键点标注信息文件(xxxx.txt)，放置于 ./datasets/annotations/下（train存放训练图片的标注文件，val存放验证图片的标注文件），存放目录结构如下：
+
+```
 ├── annotations
-│   ├── train_wider_face.json
-│   ├── val_wider_face.json
+│   ├── train
+│       ├── xxx.txt
+│   ├── val
+│       ├── xxx.txt
+```
 
+特别地，标注信息的格式为：
+
+```
+# img_file/image_name # #+空格+img_file/image_name
+x, y, w, h, left_eye_x, left_eye_y, flag, right_eye_x, right_eye_y, flag, nose_x, nose_y, flag, left_mouth_x, left_mouth_y, flag, right_mouth_x, right_mouth_y, flag, confidence  # x和y是检测框左上角的坐标
+```
+
+举个例子：
+
+```
+# 0--Parade/0_Parade_marchingband_1_849.jpg
+449 330 122 149 488.906 373.643 0.0 542.089 376.442 0.0 515.031 412.83 0.0 485.174 425.893 0.0 538.357 431.491 0.0 0.82
 ```
 
+3. 生成训练所需的json格式标注数据：
+
+```
+cd ./datasets
+python gen_data.py --mode train
+```
+
+执行完成后会在./datasets/labels下生成训练数据的标注文件 train_face.json
+
+
 ## 训练
 ### 单机单卡
 
@@ -110,7 +162,7 @@ python test_wider_face.py
 ### 精度
 WIDER_FACE验证集上的测试结果如下
 
-| Method | Easy | Medium | Hard|
+| Method | Easy(p) | Medium(p) | Hard(p)|
 |:--------:| :--------:| :---------:| :------:|
 | ours(one scale) | 0.9264 | 0.9133 | 0.7479 |
 | original | 0.922 | 0.911  | 0.782|

datasets/gen_data.py

+# -*- coding:utf-8 -*-
+
+"""只需要按照实际改写images/annotations/categories另外两个字段其实可以忽略
+在keypoints/categories内容是固定的不需修改
+"""
+
+import os
+import json
+import re
+import cv2
+import argparse
+
+from tqdm import tqdm
+
+import numpy as np
+
+
+def get_landmark(txt_path, save_path):
+    txt_write = open(save_path, 'w')
+    annotationfile = open(txt_path)
+    min_bbox = 10
+    blur_value = 0.3
+    Init = False
+    img_path = None
+    bbox_landmarks = []
+    while (True):
+        line = annotationfile.readline()[:-1]
+        if not line:
+            break
+        if re.search('jpg', line):
+            if Init:
+                if len(bbox_landmarks) < 1:
+                    continue
+                txt_write.write(img_path + '\n')
+                txt_write.write(str(len(bbox_landmarks)) + '\n')
+                for lm in bbox_landmarks:
+                    txt_write.write(str(lm)+'\n')
+            Init = True
+            img_path = line.split('# ')[1]  # line[2:]
+            bbox_landmarks = []
+            continue
+        else:
+            values = line.strip().split()
+            bbox = values[:4]
+            if min(int(bbox[2]), int(bbox[3])) < min_bbox:
+                continue
+            if len(values) > 4:
+                if float(values[19]) < blur_value:
+                    continue
+                for li in range(5):
+                    value = float(values[(li+2)*3])
+                    if value == 0:  # visible
+                        values[(li + 2) * 3] = str(2)
+                    elif value == 1:  # visible
+                        values[(li + 2) * 3] = str(1)
+                    else:
+                        values[3*li+4] = str(0)
+                        values[3*li+5] = str(0)
+                        values[3*li+6] = str(0)
+            values = ' '.join(values)
+            bbox_landmarks.append(values)
+    txt_write.close()
+    annotationfile.close()
+
+
+class COCO(object):
+
+    def info(self):
+        return {"version": "1.0",
+                "year": 2020,
+                "contributor": "Mr.yang",
+                "date_created": "2018/08/21",
+                "github": "https://github.com/bleakie"}
+
+    def licenses(self):
+        return [
+            {
+                "url": "http://creativecommons.org/licenses/by-nc-sa/2.0/",
+                "name": "Attribution-NonCommercial-ShareAlike License",
+                "id": 1
+            }
+        ]
+
+    def image(self):
+        return {
+            "license": 4,
+            "file_name": "000000397133.jpg",  # 图片名
+            "coco_url":  "http://images.cocodataset.org/val2017/000000397133.jpg",  # 网路地址路径
+            "height": 427,  # 高
+            "width": 640,  # 宽
+            "date_captured": "2013-11-14 17:02:52",  # 数据获取日期
+            "flickr_url": "http://farm7.staticflickr.com/6116/6255196340_da26cf2c9e_z.jpg",  # flickr网路地址
+            "id": 397133  # 图片的ID编号（每张图片ID是唯一的）
+        }
+
+    def annotation(self):
+        return {
+            "segmentation": [  # 对象的边界点（边界多边形）
+                [
+                    0., 0.,  # 第一个点 x,y坐标
+                    0., 0.,  # 第二个点 x,y坐标
+                    0., 0.,
+                    0., 0.
+                ]
+            ],
+            "num_keypoints": 5,
+            # keypoints是按照以下关键点来标记的，如果nose 没有标则为0,0,0(3个数字为一组，分别为x,y,v v=0表示为标记此时的x=y=0,
+            # v=1表示标记了但是在图上是不可见，v=2表示标记了，在图上可见)
+            "keypoints": [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+            "area": 0.,  # 区域面积
+            "iscrowd": 0,
+            "image_id": 397133,  # 对应的图片ID（与images中的ID对应）
+            "bbox": [0., 0., 0., 0.],  # 定位边框 [x,y,w,h]
+            "category_id": 1,  # 类别ID（与categories中的ID对应）
+            "id": 82445  # 对象ID，因为每一个图像有不止一个对象，所以要对每一个对象编号（每个对象的ID是唯一的）
+        }
+
+    def categorie(self):
+        return {
+            "supercategory": "face",  # 主类别
+            "id": 1,  # 类对应的id （0 默认为背景）
+            "name": "face",  # 子类别
+            "keypoints": ["left_eye", "right_eye", "nose", "left_mouth", "right_mouth"],
+            # "skeleton": [[1, 3], [2, 3], [3, 4], [3, 5]]
+        }
+
+
+class Keypoints2COCO(COCO):
+    def __init__(self, label_path, save_json_path, images_path):
+        self.label = open(label_path, )
+        self.save_json_path = save_json_path  # 最终保存的json文件
+        self.images_path = images_path  # 原始图片保存的位置
+        self.images = []
+        self.annotations = []
+        # self.label = []
+        self.annID = 1
+        self.height = 0
+        self.width = 0
+        self.num = 1
+        self.keypoints = ["left_eye", "right_eye",
+                          "nose", "left_mouth", "right_mouth"]
+        self.num_keypoints = 5
+
+    def __call__(self):
+        while (True):
+            img_path = self.label.readline()[:-1]
+            if not img_path:
+                break
+            if img_path.endswith('.jpg'):
+                img_full_path = os.path.join(self.images_path, img_path)
+                if not os.path.exists(img_full_path):
+                    # print("img not exist", img_full_path)
+                    continue
+                # print("img_full_path", img_full_path)
+                # init image
+                image = self.image()
+                image["file_name"] = img_path
+                image["id"] = self.num
+                img = cv2.imread(img_full_path)
+                if img is None:
+                    continue
+
+                image["height"] = img.shape[0]
+                image["width"] = img.shape[1]
+
+                line = self.label.readline()[:-1]
+
+                if not line:
+                    break
+                facenum = (int)(line)
+                # print("facenum", facenum)
+                # init annotation
+                annotation = self.annotation()
+                for _ in range(facenum):
+                    line = [float(x)
+                            for x in self.label.readline().strip().split()]
+                    # print("***", line)
+                    bbox = list(line[:4])
+                    if len(line) > 4:
+                        line[6], line[9], line[12], line[15], line[18] = int(line[6]), int(
+                            line[9]), int(line[12]), int(line[15]), int(line[18])
+                        index = [line[6], line[9],
+                                 line[12], line[15], line[18]]
+                        self.num_keypoints = len(np.minimum(index, 1))
+                        annotation['keypoints'] = line[4:-1]  # 默认为可见 v=2
+                        annotation['num_keypoints'] = self.num_keypoints
+                    annotation["image_id"] = self.num
+                    annotation["id"] = self.annID
+                    annotation["bbox"] = bbox
+                    annotation['area'] = bbox[2]*bbox[3]
+                    annotation['segmentation'] = [
+                        bbox[0], bbox[1], bbox[0]+bbox[2], bbox[1]+bbox[3]]
+                    self.annotations.append(annotation)
+
+                    self.annID += 1  # 对应对象
+                    annotation = self.annotation()  # 计算下一个对象
+
+                self.num += 1  # 对应图像
+                self.images.append(image)
+
+        jsdata = {"info": self.info(), "licenses": self.licenses(), "images": self.images,
+                  "annotations": self.annotations, "categories": [self.categorie()]}
+
+        json.dump(jsdata, open(self.save_json_path, 'w'), indent=4,
+                  default=float)  # python3 需加上default=float 否则会报错
+
+
+parser = argparse.ArgumentParser()
+# basic experiment setting
+parser.add_argument('--mode', default='train',
+                    help='Please input train or val')
+
+opt = parser.parse_args()
+
+if __name__ == "__main__":
+    ROOT_PATH = './'
+    label_type = opt.mode
+    img_path = os.path.join(ROOT_PATH, 'images', f'{label_type}')
+    txt_path = os.path.join(ROOT_PATH, 'annotations', f'{label_type}/{label_type}.txt')
+    save_landmark_path = txt_path[:txt_path.rfind('/')]
+    save_landmark_path = os.path.join(
+        save_landmark_path, f'landmark_{label_type}.txt')
+    get_landmark(txt_path, save_landmark_path)  # 处理关键点标注数据
+
+    save_label_path = os.path.join(ROOT_PATH, 'labels')
+    if not os.path.exists(save_label_path):
+        os.makedirs(save_label_path)
+    save_label_path = os.path.join(
+        save_label_path, f'./{label_type}_face.json')
+    Keypoints2COCO(save_landmark_path, save_label_path, img_path)()
+    print("dealing labels end.")