Fangchang Ma

I think you should avoid converting rgb images (uint8) to float. If you really need to do that, scale the color to between [0,1] after the float-type conversion. Please let...

Hi @godman2016 , `nValidElement`, computed as the number of non-zero points in the target image, is in fact the number of lidar measurements on the image plane.

> Thks for sharing your perfect job. > I'm interested in your dense-map from VOSLAM. But I don't know how to generate sparse depth image. > Can you provide your...

Hey @TruongKhang. You were right - the exact sampling pattern changes from frame to frame, and from epoch to epoch. This randomization is by design, such that the network does...

> is it reasonable when compared with the baseline methods Which baseline methods were you referring to?

That's indeed a problem. However, `work_id+args.epochs` is probably not the ideal solution because of possible ID overlapping. For instance, work_id=20 at epochs=15 overlaps with work_id=15 at epochs=20 when resuming.

Preferably we don't have to rebuild the data loaders at every epoch. Additionally, this workaround might still suffer from seed overlapping if one resumes with a different number of workers.

The `.tar` files are not compressed - they are simply pytorch-saved dictionaries. In other words, load them directly from pytorch without any decompression.

Thanks for the pull request. To gain a better understanding of the PR: where exactly in the code does the "overriding" create a problem (in the DepthCompletionNet definition, or in...

You were correct about the rightmost column being the semi-dense annotations (i.e., `gt` in the following line). https://github.com/fangchangma/self-supervised-depth-completion/blob/b8140669c2ff7658509edd39f5a73638b78d214b/main.py#L101 The fourth column is currently part of the visualization, regardless of how...