dwhou / emdc-pytorch

PyTorch version of the paper 'Learning an Efficient Multimodal Depth Completion Model' (ECCVW 2022)

Hi there 👋

emdc-pytorch's Issues

Request for ground truth of testing data which used in RGB+TOF Depth Completion @MIPI-challenge

Hi @dwHou,

Thanks for your great work! I'm curious about there is anyway to get ground truth(GT ) of testing data(not the validation_input)?

The reason I'm pursuing GT is because I want to reproduce the performance you mentioned in your paper. However, the GT depth are not available to participants in the challenge, the official said.

About the license for this model

Thank you for sharing your great code. 😺

What is the license for this model? I'd like to cite it to the repository I'm working on if possible, but I want to post the license correctly.
https://github.com/PINTO0309/PINTO_model_zoo

Thank you.

how to test on custom images?

divide loss_sam1 by (loss_sam1/loss1).detach()

Hi @dwHou ,

Thanks for your great work again! One more question, I'm curious about why you divide loss_sam1 by (loss_sam1/loss1).detach()

Does it equal to loss = loss + loss_sam1 * (loss1/loss_sam1) in math? and why?

EMDC-PyTorch/train_model.py

Lines 127 to 133 in 07867c7

	if 'criterion1' in args.loss:
	loss1 = criterion1(output, target)
	loss = loss1
	loss_sam1 = criterion1(y_loc, target)
	loss += loss_sam1/(loss_sam1/loss1).detach()
	loss_sam2 = criterion1(y_glb, target)
	loss += loss_sam2/(loss_sam2/loss1).detach()

Looking forward to your reply, thanks.

have you done experiments on the NYU or KITTI depth completion datasets?

I am very interested in the FCSPN structure you proposed, and want to compare this method with several mainstream depth completion methods. So please tell me if you have done experiments on the NYU or KITTI dataset

我对您提出的FCSPN结构非常感兴趣，想对比一下这个方法和几个主流深度补全方法。所以请教一下您是否在NYU或者KITTI数据集上做过实验

想问下这个深度填充对人像的深度补充有泛化性吗

您好，论文中所展示的深度填充大多为场景，想问下该模型对人物深度填充有泛化性吗？还是需要自己重新训练？

Model does not converge

Thank you so much for sharing.
Unfortunately, the model did not converge though I trained the model using "bash train.sh" on the dataset you offered with us. The only change I made was just the data directory. Could you kindly tell me how you set up those hyper-parameters? Thank you in advance.