Questions about mmpose.evaluation.metrics.CocoMetric

[WinnerMeat]

In the configuration file of YOLOx-Pose, we need to set the content of val_evaluator, in which val_evaluator.score_mode has three optional parameters: bbox, bbox_keypoint and bbox_rle, in which bbox means that the score of the bbox will be used as the prediction result.
Is it understood that when val_evaluator.score_mode=bbox, the detection accuracy obtained by the program is the accuracy about the bounding box; val_evaluator.score_mode=bbox_keypoint, the detection accuracy obtained by the program is the accuracy about the keypoint of the human body.
The example of val_evaluator in the configuration file is as follows：
val_evaluator = dict(
type='mmpose.CocoMetric',
ann_file='***.json',
score_mode='bbox_keypoint',
nms_mode='none',
use_area=False,
iou_type='keypoints_crowd',
prefix='crowdpose'
)
If anyone knows, I would appreciate an answer as soon as possible, thanks！

[Ben-Louis]

This parameter specifies how the confidence score of each prediction is calculated. For more information, you can refer to

mmpose/mmpose/evaluation/metrics/coco_metric.py

Lines 453 to 475 in 5a3be94

	if self.score_mode == 'bbox':
	instance['score'] = instance['bbox_score']
	elif self.score_mode == 'keypoint':
	instance['score'] = np.mean(instance['keypoint_scores'])
	else:
	bbox_score = instance['bbox_score']
	if self.score_mode == 'bbox_rle':
	keypoint_scores = instance['keypoint_scores']
	instance['score'] = float(bbox_score +
	np.mean(keypoint_scores) +
	np.max(keypoint_scores))
	else: # self.score_mode == 'bbox_keypoint':
	mean_kpt_score = 0
	valid_num = 0
	for kpt_idx in range(num_keypoints):
	kpt_score = instance['keypoint_scores'][kpt_idx]
	if kpt_score > self.keypoint_score_thr:
	mean_kpt_score += kpt_score
	valid_num += 1
	if valid_num != 0:
	mean_kpt_score /= valid_num
	instance['score'] = bbox_score * mean_kpt_score

. In CocoMetric, the confidence score affects the order of predictions and, as a result, influences AP.

[WinnerMeat]

Thanks for the reply.