Does segmentation branch have a significant impact on performance
Opened this issue · 0 comments
2033329616 commented
I found a segmentation branch in the code, but in the paper there is no segmentation, and whether repeated experiments can achieve the same effect in paper if I remove segmentation branch?
loss = cls_loss * cls_loss_weight + reg_xytl_loss * xyt_loss_weight \
+ seg_loss * seg_loss_weight + iou_loss * iou_loss_weight
return_value = {
'loss': loss,
'loss_stats': {
'loss': loss,
'cls_loss': cls_loss * cls_loss_weight,
'reg_xytl_loss': reg_xytl_loss * xyt_loss_weight,
'seg_loss': seg_loss * seg_loss_weight,
'iou_loss': iou_loss * iou_loss_weight
}
}