Self-supervised-Monocular-Trained-Depth-Estimation-using-Self-attention-and-Discrete-Disparity-Volum icon indicating copy to clipboard operation
Self-supervised-Monocular-Trained-Depth-Estimation-using-Self-attention-and-Discrete-Disparity-Volum copied to clipboard

Published 20 hours ago • verified publisher icon sjsu-smart-lab

the 'holes' of infinits depth

Open kamiLight opened this issue 4 years ago • 0 comments

Hi! Thanks for your excellent work. will you release the pretrained model in the future work? Recently, i train my own model resnet101 w/ dilated convolution, using self.encoder = resnets[num_layers](replace_stride_with_dilation=[False, True, True], pretrained=pretrained), encoder will output a tensor of size N x H/8 x W/8. Just like your work with ResNet101 w/ Dilated Conv, although exists a little difference. And i encounter a problem that, as shown in figure，the 'holes' of infinits depth appears in predicted depth maps, for objects that are typically observed to be moving during training. This issue has been solved in monodepth2 with auto-masking, but the masks seems to be disabled when using resnet101 w/ dilated convolution. So i hope to know whether you have same issue in your experiments? Thanks!

7_disp

kamiLight avatar Mar 15 '21 01:03 kamiLight