search

ilovepose / DarkPose

Distribution-Aware Coordinate Representation for Human Pose Estimation
https://ilovepose.github.io/coco
Apache License 2.0
550 stars 80 forks source link

Why use different BLUR_KERNEL in post-processing? #22

Closed rhifbjbfdg closed 3 years ago

rhifbjbfdg commented 3 years ago

For the input size of 128x96, BLUR_KERNEL is 3. However, BLUR_KERNEL is 11 for the input size of 256x192. Can you explain the reason? Thanks.

xizero00 commented 3 years ago

Because the input size is increased. In fact, for the large input size,
if you set the kernel size to 3 or 7 or 9 or 11, you can obtain almost the same AP results (although the AP is increased slightly). For small input size, the kernel size has a significant impact on the performance, you should set it to a small value.