search

maudzung / TTNet-Real-time-Analysis-System-for-Table-Tennis-Pytorch

Unofficial implementation of "TTNet: Real-time temporal and spatial video analysis of table tennis" (CVPR 2020)
https://arxiv.org/pdf/2004.09927.pdf
600 stars 158 forks source link

Not producing comparable results for ball detection and event classification modules. #43

Closed AugustRushG closed 1 month ago

AugustRushG commented 1 month ago

Hi, I'm just wondering has anybody got comparable results with the paper using this code? I have trained it using what is provided but the RMSE is way too large. In their paper, they stated that their RMSE for global is 3.60 and for local is 1.46. But what I got after training RMSE for global is 46.8 and for local is 90.5, there is quite a big margin between. This happened to the event classification module as well, where the PCE is supposed to be 0.979 and SPCE is 0.975 but I am getting PCE 0.582 and SPCE 0.838.

Also, there is some mismatch between this repo and the paper on the implementation of the ball detection module. In paper image there are two FC layers after FC2 one produces X and one produces Y. This repo only has one FC which produces a vector with length 448, it is the width+height (320,128). I modified it and got the right result.

Just wondering has anybody got comparable results with the ball detection module?

AugustRushG commented 1 month ago

Hi guys, I have fixed all problems and also modified the training code to produce good results, please refer to https://github.com/AugustRushG/TTNet-Real-time-Analysis-System-for-Table-Tennis-Pytorch.