Looking forward to the code of TokenPose version

[yshMars]

Hi, It's an amazing work! I reproduce the code of SimDR-Tokenpose version based on mmpose, but didn't get a satisfying performance (Actually a very low performance, somewhere goes wrong definitely...). I'm looking forward to the open source of SimDR - Tokenpose version. Thanks a lot !!

[yshMars]

I reproduce Tokenpose on mmpose and it works fine, but when I add SimDR and something goes wrong. The training loss decreases well and performance AR is high, but mAP is very low (based on COCO). I checked every code I modified and I don't know where goes wrong. T T

[yshMars]

I think I know where went wrong... You use the max value of max_val_x or max_val_y to be the 'max_vals' while inference, but actually after softmax the max value might be very small cause the distribution of output is not so centralized, there might be noise. After I force the max value to 1, I got a good performance. Sorry for disturbing.

[DaMiBear]

@yshMars Hi, Do you mean that you filter the confidence(such as max value >0.3) when evaluating ? The max value may only be 0.1, which causes this keypoint to be ignored after filtering？