feihuzhang / dsmnet Goto Github PK

Hello, In the paper you mentioned that

For 3D cost aggregation of the cost volume, two non-local filters are further added for cost volume filtering in each channel/depth.

I don't understand this sentence so I looked up your code, but you didn't add NLFilter3D Module into the cost aggregation network.
From the code I guess you mean that it does NLF along the Feature channel and the Disparity channel to get 2 tensors, and the concat them to do 3D conv. Is that right?

I want to finetune the network with my own data set. How do I generate JSON files?

Thank you very much again for your sharing. May I venture to ask if the new supplementary synthetic dataset can be Shared?

@feihuzhang ,Thank you for the code and paper. When I read the paper, it said that domain-invariant can be easily extended to the field of optical flow.
But I want to ask, for example, which part of PWC-Net should the Domain Norm be added to? (just like Domain Norm is added to the feature extraction and Guidance branch of DSMNet)

Where is the file?

Your paper is good.

Hi,

May I know if you have deleted your submission to KITTI15 benchmark? I can't find your results in the official benchmarks.

Would like to use some images of your results from the benchmark for comparisons.

Cheers,
Sam

I can successfully conduct compile.sh, and **Already get libs/GANet/build/lib/GANet.cpython-37m-x86_64-linux-gnu.so file.**But when i conduct predict.sh, it report an error:
File "/home/common/xxx/DSMNet/libs/GANet/functions/GANet.py", line 3, in <module> from ..build.lib import GANet ImportError: cannot import name 'GANet'

Hello, we want to know how do you select the KITTI, Middlebury and ETH3D validation datasets in table 3 so that we can make a direct comparision in our paper. Do you just use all the training images in KITTI, Middlebury and ETH3D?

Hi,

Thank you for the great and interesting work. Looking forward to the release of your pre-trained weights and additional synthetic data.

I have a question regarding the feature maps presented in Figure 1 in your paper. It highlights the important features that are extracted by the model at an early stage. May I know how did you generate those images?

Thank you 👍🏻

Hello, @feihuzhang thanks for your work.
However, your codes have some problems. I can not run this project correctly. Can you give me the correct code that you have run? And the model have to be trained for long time, can you give me the final model to predict? Thank you!

Hello:

Thanks so much for open-sourcing the code. The presentation also looks nice!

I wonder what settings did you use for your reported result for SGM, because there are lots of parameters one can tune. Is this an already implemented version from libraries such as OpenCV, or is it a customized version? Do you mind sharing more details? Thank you!

@feihuzhang Thank you for the code and paper. Can you please share information about the speed of the model, and the device the model is tested on?
I am also looking forward to the pretrained model.

Thanks for your amazing work. I wonder when will calar dataset be published ?

Hello, I read your paper, and get a little a little confused about Figure 3 in it. What does the y axis (1,2,3,4,5) of Figure 3 refer to? After Domain transform,the norm of the C-channels feature is not 1(before scale)?