Hi,

Thank you. I have a question on equation 5 and 6 of the paper. In those equations, since both depths of i and j are normalized, they both become 1 or not ?

Hi @hyBlue ,

Thank you for your great work. I just have one question regarding to the GT depth used for evaluation. In your dataloader, you load gt_depth from file. (https://github.com/hyBlue/FSRE-Depth/blob/d746866965204c37a072f3518dfd873df7ed2f1a/datasets/kitti_dataset.py#L84). However, in your evaluation code, gt_depth is loaded from an npz file. (https://github.com/hyBlue/FSRE-Depth/blob/d746866965204c37a072f3518dfd873df7ed2f1a/evaluate_depth.py#L148) And the gt_depth from dataloader is not used?

Is this true?

Also, for the gt_depth from npz, this gt_depth is corresponding to the groundtruth set of KITTI Depth Prediction or velodyne_raw set of KITTI Depth Prediction?

Thank you in advance.

Best,

Hi, the labels I got using NIVDA's segmentation network are pseudo-labels for 19 categories, which are not the same as the pre-computed segmentation you uploaded

Could you please share your steps and practices?

Thank you very much!

Hi, I used your architecture to train my own dataset, but the results are not as expected, the evaluation results and images are quite poor.

Illogical oddities occur when inferring images. I see that you do not have the inference part.

Could you please give me some advice? Thank you !

Hi,

I try to train and test your model. I found some error in your code.

1. '--data_path' option -> when I set --data_path for my own dataset path, the path that your code modified was pointed to wrong directory
2. when I run 'evaluate_depth.py' for tersting, 2-1. '--fix_K' option does not exist in KittiDataset class 2-2. in 111 line, output = models['depth'](features, mode='depth'), return error => forward() got an unexpected keyword argument 'mode'

I hope to use your official evaluate code / script.

Thanks!

Hi @hyBlue,

What method did you use to handle discoutinuity of GT depth when plotting the error map in your paper? Did you use interpolation? Since GT depth map inevitably has empty pixel, I am just wondering about this. In addition, I think there is a bug in your evaluation code in your current implementation.

https://github.com/hyBlue/FSRE-Depth/blob/d746866965204c37a072f3518dfd873df7ed2f1a/evaluate_depth.py#L68 the dataset doesn't have input: fix_K.

So, I am looking forward to your reply on error map part. Thank you in advance.

Best,

when i run evaluate_depth /home/shs/miniconda3/envs/shs/lib/python3.7/site-packages/torchvision/transforms/transforms.py:334: UserWarning: Argument interpolation should be of type InterpolationMode instead of int. Please, use InterpolationMode enum. "Argument interpolation should be of type InterpolationMode instead of int. "

Hi Thank you for a great job. Could you tell me how the sgementation dataset is obtained? Is it used to train semantic segmentation subnetworks ？ Thank you very much.

Hi, thanks for your good work. I have some confusion, hope for your answer. when calculating Cross-task Similarity, you use att = (q_out * k_out).sum(dim=2) / np.sqrt(self.out_channel) tensor mul is that (B, Head, C, H, W) * (B, Head, C, H, W) = B, Head, C, H, W I think Cross-task Similarity is like cross-task self-attention, is it right? Usually, self-attention is commmon that (B, Head, C, HW) * (B, Head, HW, C) = B, Head, HW* HW my qusetion: what is the difference between Cross-task Similarity and self-attention? Or Cross-task Similarity (B, Head, C, H, W) * (B, Head, C, H, W) = B, Head, C, H, W is a special way to calculate self-attention?