使用txt格式的ground truth文件测人脸检测的AP

使用txt格式的ground truth文件测人脸检测的AP

使用txt格式的ground truth文件测人脸检测的AP

• • 起源
  • 代码
  • GitHub地址

起源

widerface官方的ground truth文件是mat格式的，因为要测试自己的数据，gt文件是txt格式的，所以要修改一下测AP的方法。

代码

测AP的原代码中的部分函数可以直接使用，修改读gt文件和获取gt中人脸框的部分即可，修改部分如下

def evatxt(pred, gt_file, iou_thresh):pred = get_preds(pred)norm_score(pred)boxes_list = get_gt_boxes_from_txt(gt_file)thresh_num = 1000aps = []count_face = 0pr_curve = np.zeros((thresh_num, 2)).astype('float')  # control calcultate how many sampleserror_count = 0for boxes in boxes_list:try:image_name = boxesevent_name = image_name.split('/')[0]image_name = image_name.split('/')[1]name_all = event_name + '\' + image_name[:-4]pred_list = pred[event_name]pred_info = pred_list[name_all]except:error_count += 1continuegt_boxes = boxes_list[boxes].astype('float')if len(pred_info) == 0 or len(gt_boxes) == 0:error_count += 1continuecount_face += len(gt_boxes)pred_recall, proposal_list = image_eval(pred_info, gt_boxes, iou_thresh)_img_pr_info = img_pr_info(thresh_num, pred_info, proposal_list, pred_recall)pr_curve += _img_pr_infoprint(boxes)pr_curve = dataset_pr_info(thresh_num, pr_curve, count_face)propose = pr_curve[:, 0]recall = pr_curve[:, 1]ap = voc_ap(recall, propose)aps.append(ap)print("==================== Results ====================")print("Val AP: {}".format(aps[0]))print("=================================================")

GitHub地址