search

IDEA-Research / ED-Pose

[ICLR 2023] Official implementation of the paper "Explicit Box Detection Unifies End-to-End Multi-Person Pose Estimation "
Other
153 stars 10 forks source link

Question about the reported metrics of Poseur in your paper #25

Open Kelly510 opened 9 months ago

Kelly510 commented 9 months ago

Hi. I have a question about the reported metrics of Poseur in your paper. In the original Poseur paper, AP is 75.4 with ResNet-50 backbone. However, it is 74.2 in your paper with the same backbone. May I ask the reason of this difference?

image 1703596696096

yangjie-cv commented 9 months ago

In the original Poseur paper, an AP of 75.4 was achieved through the use of the flip test as a test-time augmentation technique. However, in the case of ED-Pose, no test-time augmentation is employed. We conduct a fair comparison between the two methods, where we utilize the Poseur checkpoints while excluding the flip test and yield an AP of 74.2.

Kelly510 commented 9 months ago

In the original Poseur paper, an AP of 75.4 was achieved through the use of the flip test as a test-time augmentation technique. However, in the case of ED-Pose, no test-time augmentation is employed. We conduct a fair comparison between the two methods, where we utilize the Poseur checkpoints while excluding the flip test and yield an AP of 74.2.

I got it. Thanks for your reply very much!