how to implement softer-NMS by pytorch version

[gittigxuy]

just change a little bit code in soft NMS ？could you please illustrate detailly?

[LiDaiY]

I saw the source code，it seems we should predict the standard variance before executing NMS. So there should be more change?

[JoyHuYY1412]

I implemented only the KL-loss part in pytorch.

I haven't test but you can check dies it work. The bbox_pred_std is the 'alpha' in the paper.

  def smooth_l1_loss(self, bbox_pred, bbox_targets, bbox_inside_weights, bbox_outside_weights, sigma=1.0):    
       sigma_2 = sigma ** 2
       box_diff = bbox_pred - bbox_targets
       in_box_diff = bbox_inside_weights * box_diff
       abs_in_box_diff = torch.abs(in_box_diff)
       smoothL1_sign = (abs_in_box_diff < 1. / sigma_2).detach().float()
       loss_box = (torch.pow(in_box_diff, 2) * (sigma_2 / 2.) * smoothL1_sign
                   + (abs_in_box_diff - (0.5 / sigma_2)) * (1. - smoothL1_sign)) * bbox_outside_weights
       return loss_box.sum() / loss_box.shape[0]

   def KL_loss(self, bbox_pred, bbox_targets, bbox_pred_std, bbox_inside_weights, bbox_outside_weights, sigma=1.0):
       #KL-loss
       sigma_2 = sigma ** 2
       box_diff = bbox_pred - bbox_targets
       in_box_diff = bbox_inside_weights * box_diff #bbox_inw = in_box_diff
       bbox_l1abs = torch.abs(in_box_diff)  #abs_in_box_diff = bbox_l1abs
       # bbox_sq = in_box_diff * in_box_diff
       smoothL1_sign = (bbox_l1abs < 1. / sigma_2).detach().float()  #1 if bbox_l1abs<1 else 0 
       bbox_inws = (torch.pow(in_box_diff, 2) * (sigma_2 / 2.) * smoothL1_sign
                   + (bbox_l1abs - (0.5 / sigma_2)) * (1. - smoothL1_sign)) 
       bbox_inws = bbox_inws.detach().float()  #?? to be confirmed
       scale = 1
       bbox_pred_std_abs_log = bbox_pred_std*0.5*scale
       bbox_pred_std_nabs = -1*bbox_pred_std
       bbox_pred_std_nexp = torch.exp(bbox_pred_std_nabs)
       bbox_inws_out = bbox_pred_std_nexp * bbox_inws
       bbox_pred_std_abs_logw = bbox_pred_std_abs_log * bbox_outside_weights
       bbox_pred_std_abs_logwr = torch.mean(bbox_pred_std_abs_logw, dim = 0)
       
       #bbox_pred grad, stop std
       loss_bbox = self.smooth_l1_loss(bbox_pred, bbox_targets, bbox_inside_weights, bbox_pred_std_nexp)
       bbox_pred_std_abs_logw_loss = torch.sum(bbox_pred_std_abs_logwr)
       bbox_inws_out = bbox_inws_out * scale
       bbox_inws_outr = torch.mean(bbox_inws_out, dim = 0)
       bbox_pred_std_abs_mulw_loss = torch.sum(bbox_inws_outr)
       return (loss_bbox + bbox_pred_std_abs_logw_loss + bbox_pred_std_abs_mulw_loss)

[wysot]

@JoyHuYY1412
What are the bbox_inside_weights and bbox_outside_weights in the KL_LOSS function?

[ethanhe42]

3rd party reimplementation in PyTorch: ethanhe42/KL-Loss#20
BTW, this repo is deprecated. Please use our new repo: https://github.com/yihui-he/KL-Loss