从wider_attribute数据集中提取性别

本文提供了一个从wider_attribute数据集中提取性别属性的方法；并提供了数据集的下载链接。

目录

1、数据集说明 

2、方法

3、附

4. Citation

---

1、数据集说明 

WIDER Attribute是一个大规模的人类属性数据集。它包含属于 30 个场景类别的 13789 个图像，以及 57524 个人体边界框，每个边界框都用 14 个二进制属性进行注释。

图片引自： Human Attribute Recognition by Deep Hierarchical Contexts

        数据集一览：

 

 

        可见，该数据集也是从大图里面标注了人体的边框，并提供了每个人的属性集的。

        此外，该数据集相比之前的其他数据集也有一定优势：

        数据标注格式为：

{
	"images"		: [image],			# A list of "image" (see below)
	"attribute_id_map"	: {str : str},			# A dict, mapping attribute_id to attribute_name
	"scene_id_map"		: {str : str}			# A dict, mapping scene_id to scene_name
}

image {
	"targets"		: [target],			# A list of "target" (see below)
	"file_name"		: str,				# Image file name
	"scene_id"		: int				# Scene id
}

target {
	"attribute"		: [int]				# A list of int, the i-th element corresponds to the i-th attribute, and the value could be 1(possitive), -1(negative) or 0 (unspecified)
	"bbox"			: [x, y, width, height] 	# Human bounding box
}

2、方法

通过以下代码，可以从数据集中解析出每个图片的属性，并根据性别存放到不同文件夹：

"""
parse person's gender from wider_attribute
"""
import os
import cv2
import glob
import json
import numpy as np


def xywh2xyxy(xywh):
    x1, y1, w, h = xywh
    x2 = x1 + w
    y2 = y1 + h
    return [int(x1), int(y1), int(x2), int(y2)]


def main(raw_path, new_dataset_path):
    classes = ['0_Female', '1_Male']
    os.makedirs(new_dataset_path, exist_ok=True)
    Male_path = os.path.join(new_dataset_path, '1_Male')
    Female_path = os.path.join(new_dataset_path, '0_Female')
    os.makedirs(Male_path, exist_ok=True)
    os.makedirs(Female_path, exist_ok=True)

    anno_files = glob.glob(raw_path + '/annotation/*.json')
    for anno_file in anno_files:
        print("\n\nProcessing: {}".format(anno_file))
        anno = json.load(open(anno_file, 'r'))
        images = anno['images']
        attribute_id_map = anno['attribute_id_map']

        imgs_num = len(images)
        for idx, image in enumerate(images):
            print("\rdeal: [{}/{}]".format(idx + 1, imgs_num), end='')
            file_name = image['file_name']
            img_cv = cv2.imread(os.path.join(raw_path, 'Image', file_name))
            H, W, _ = img_cv.shape
            targets = image['targets']

            for t_i, target in enumerate(targets):
                attribute = target['attribute']
                if attribute[0] == 0:
                    continue
                cls_id = 1 if attribute[0] == 1 else 0
                class_name = classes[cls_id]
                bbox = target['bbox']
                x1, y1, x2, y2 = xywh2xyxy(bbox)
                x1, x2 = np.clip([x1, x2], 0, W)
                y1, y2 = np.clip([y1, y2], 0, H)
                if y2-y1 <= 0 or x2-x1 <= 0:
                    continue
                target_cv = img_cv[y1:y2, x1:x2, :]
                # print(target_cv.shape, [x1, y1, x2, y2])

                name = os.path.splitext(file_name.split('/')[-1])[0] + '_{}.jpg'.format(t_i)
                img_save_to = os.path.join(new_dataset_path, class_name, 'wider_' + name)
                write_status = cv2.imwrite(img_save_to, target_cv)
                assert write_status, "target image [{}] write error!".format(img_save_to)


if __name__ == '__main__':
    raw_path = '../wider_attribute'
    new_path = '../datasets/gender_dataset_parse/from_wider'
    main(raw_path, new_path)

3、附

下载链接: https://pan.baidu.com/s/1mbxI6npcxsx0FHQh2DV_OQ?pwd=b1xw 提取码: b1xw 

4. Citation

@inproceedings{li2016human,
  author = {Li, Yining and Huang, Chen and Loy, Chen Change and Tang, Xiaoou},
  title = {Human Attribute Recognition by Deep Hierarchical Contexts},
  booktitle = {European Conference on Computer Vision},
  year = {2016}
}

论文：http://personal.ie.cuhk.edu.hk/~ccloy/files/eccv_2016_human.pdf