Preciosn Recall for bbox or keypoint?

[CyberJZTXJ]

Hi, I'm writing my paper and I'm using yolov8pose and I need to compare it with yolov5pose and v7pose, but I'm a little confused about the training results, here are the training results of my yolov8pose with Box and Pose both providing P and R

and here are the training results of my yolov5pose with only P and R. But it doesn't say whether it's from Box or Pose.

So I have the following questions（you may have referred to v5pose and v7pose in your design and designed it better, so I sincerely ask you questions）:

1. What is the Pose(P R) of yolov8pose and how is it calculated.
2. Is there any difference between Pose(P R) and BOX(P R) of yolov8pose？I have encountered BOX(P R) and Pose(P R) the same before.
3. what does P R mean in yolov5pose, for box or for keypoint？ to compare is it with Box (P R) or with Pose (P R)？
   Looking forward to your reply, thanks!

[pderrenger]

@CyberJZTXJ thank you for your detailed question. In YOLOv8, the precision (P) and recall (R) metrics are provided separately for both bounding boxes (Box) and keypoints (Pose). This distinction helps in evaluating the model's performance in detecting objects and their keypoints independently.

1. In YOLOv8, the Pose (P, R) metrics refer to the precision and recall of the detected keypoints. These metrics are calculated based on the accuracy of the keypoint predictions compared to the ground truth keypoints.
2. Yes, there is a difference between Pose (P, R) and Box (P, R) in YOLOv8. Box (P, R) measures the precision and recall of the bounding box detections, while Pose (P, R) measures the precision and recall of the keypoint detections. They are evaluated separately to provide a clearer understanding of the model's performance in both tasks.
3. In YOLOv5, the precision and recall metrics typically refer to the bounding box detections unless explicitly stated otherwise. For a fair comparison with YOLOv8, you should compare the Box (P, R) metrics of YOLOv5 with the Box (P, R) metrics of YOLOv8, and similarly, compare the Pose (P, R) metrics if available.

I hope this clarifies your queries. If you need further details, feel free to ask.

[CyberJZTXJ]

Thank you very much for your reply!
So what you're saying is(using v5pose as an example): I just need to compare v5pose's (mAP50 mAP50-95) to v8pose's Pose (mAP50 mAP50-95),and compare v5pose's (P R) to v8pose's Box (P R) , is that right?

[pderrenger]

@CyberJZTXJ you're welcome! To clarify, you should compare YOLOv5's mAP50 and mAP50-95 with YOLOv8's Pose mAP50 and mAP50-95 for keypoint detection performance. For precision and recall, compare YOLOv5's P and R with YOLOv8's Box P and R for bounding box detection performance. This will ensure a fair and accurate comparison between the models.

[CyberJZTXJ]

Thank you very much for your answer! For further confirmation, I put only three images in the validation set(containing three objects) and adjusted the conf of yolov5pose to 0.92 and got the following result:

This matches my results (only 2 objects were predicted)（R=2/3=0.667）
This proves that ‘the precision and recall metrics typically refer to the bounding box detections unless explicitly stated otherwise’
Your answer is very professional, once again I express my gratitude to you and wish you more and more beauty and wisdom!

[pderrenger]

Thank you for your kind words! I'm glad the clarification helped. Your validation approach and results confirm that the precision and recall metrics in YOLOv5pose indeed refer to bounding box detections. If you have any more questions or need further assistance, feel free to ask. Best of luck with your research!