深度真值使用

[suyu-star]

Thank you for your amazing project and making your code public!

  if self.depth2occ_intra:
      if self.depth_gt:
          if not self.training:
              kwargs['gt_depth']=kwargs['gt_depth'][0]
          gt_depth=self.get_downsampled_gt_depth(kwargs['gt_depth'])
              
          b,n,d,h,w=depth.shape
          gt_depth=gt_depth.reshape(b,n,h,w,d).permute(0,1,4,2,3)
          fg_mask = torch.max(gt_depth, dim=2,keepdim=True).values > 0.0
          fg_mask=fg_mask.repeat(1,1,d,1,1)
          
          depth_=depth.clone()
          depth_[fg_mask]=gt_depth[fg_mask]
      else:
          depth_=depth

the paper mentions that ground-truth depth is used to guide model training in the early stage. However, as shown in the code above, the depth used in the computation of occ_weight is expected to be the model’s own predicted depth. We find that ground-truth depth is actually used during this computation. Would this introduce unfairness or lead to an unfair comparison?

[cdb342]

Thank you for your amazing project and making your code public!

  if self.depth2occ_intra:
      if self.depth_gt:
          if not self.training:
              kwargs['gt_depth']=kwargs['gt_depth'][0]
          gt_depth=self.get_downsampled_gt_depth(kwargs['gt_depth'])
              
          b,n,d,h,w=depth.shape
          gt_depth=gt_depth.reshape(b,n,h,w,d).permute(0,1,4,2,3)
          fg_mask = torch.max(gt_depth, dim=2,keepdim=True).values > 0.0
          fg_mask=fg_mask.repeat(1,1,d,1,1)
          
          depth_=depth.clone()
          depth_[fg_mask]=gt_depth[fg_mask]
      else:
          depth_=depth

the paper mentions that ground-truth depth is used to guide model training in the early stage. However, as shown in the code above, the depth used in the computation of occ_weight is expected to be the model’s own predicted depth. We find that ground-truth depth is actually used during this computation. Would this introduce unfairness or lead to an unfair comparison?

The self.depth_gt flag is activated solely for the C+D input baseline, as shown in Tab. 1 of our paper. For camera-only configurations, this condition remains False. Additionally, the use of ground-truth depth for guiding model training is controlled by geometry_denoise.

[suyu-star]

Thank you for your amazing project and making your code public!

  if self.depth2occ_intra:
      if self.depth_gt:
          if not self.training:
              kwargs['gt_depth']=kwargs['gt_depth'][0]
          gt_depth=self.get_downsampled_gt_depth(kwargs['gt_depth'])
              
          b,n,d,h,w=depth.shape
          gt_depth=gt_depth.reshape(b,n,h,w,d).permute(0,1,4,2,3)
          fg_mask = torch.max(gt_depth, dim=2,keepdim=True).values > 0.0
          fg_mask=fg_mask.repeat(1,1,d,1,1)
          
          depth_=depth.clone()
          depth_[fg_mask]=gt_depth[fg_mask]
      else:
          depth_=depth

the paper mentions that ground-truth depth is used to guide model training in the early stage. However, as shown in the code above, the depth used in the computation of occ_weight is expected to be the model’s own predicted depth. We find that ground-truth depth is actually used during this computation. Would this introduce unfairness or lead to an unfair comparison?

The self.depth_gt flag is activated solely for the C+D input baseline, as shown in Tab. 1 of our paper. For camera-only configurations, this condition remains False. Additionally, the use of ground-truth depth for guiding model training is controlled by geometry_denoise.

好的，十分感谢您的回复