I am trying to reproduce some of the methods described in the paper with PyTorch as well.
In particular, when using Depth-error weighted (DW) SSIM and Depth consistency (DC), I observe a huge drop in the metrics at least at the beginning.
For instance, when I train monodepth or sc-sfm on 128x416 KITTI images, at the end of the first epoch, I have a log-rms of 0.27. But with DW or DC it's more in the 0.35-0.45 range, which makes me think I am doing something wrong. Or maybe it takes longer to converge. Did you also observe such things? If so, have you used some tricks to correct this?
My implementation of these methods is inspired by the TensorFlow code of the paper Unsupervised Monocular Depth Learning in Dynamic Scenes.
Hi,
I am trying to reproduce some of the methods described in the paper with PyTorch as well.
In particular, when using Depth-error weighted (DW) SSIM and Depth consistency (DC), I observe a huge drop in the metrics at least at the beginning. For instance, when I train monodepth or sc-sfm on 128x416 KITTI images, at the end of the first epoch, I have a log-rms of 0.27. But with DW or DC it's more in the 0.35-0.45 range, which makes me think I am doing something wrong. Or maybe it takes longer to converge. Did you also observe such things? If so, have you used some tricks to correct this?
My implementation of these methods is inspired by the TensorFlow code of the paper Unsupervised Monocular Depth Learning in Dynamic Scenes.
Many thanks!