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

a question about the multi-task loss function #24

Open machine981 opened 2 years ago

machine981 commented 2 years ago

In the code of the multi-task loss function, I would like to know why the classification loss is calculated in that way.

       ` total_loss += local_ball_loss / (torch.exp(2 * self.log_vars[log_vars_idx])) + self.log_vars[log_vars_idx]`

In that multi_task_loss paper (2018CVPR), author calculated classification loss through cross_entropy(CE) scaled by sigma^2 then plus log(sigma). But in your code, it seems to be calculated through CE scaled by sigma^4. I wanna know whether It's a mistake or a trick. Thanks.