longyunf / rc-pda Goto Github PK

Thanks for the wonderful work. I have one question on different levels of sparsity. If only one-pixel depth is accurate and we know the corresponding position in the image, how can this improve the overall depth prediction?

Hello,

I would like to do some tests to your work. Downloading the dataset, extracting, spliting, saving the tar in order to generate the h5 file is taking a huge amount of time and computations. Is it possible to upload the generated h5 to be able to use the model with data right away?

why would you use the cam_token2(the next token of the sample_idx) and cam_taken3(rather than cam_token and cam_token2) to calculate the image flow in prepare_flow_im.py?would you willing to share your thoughts?

Hi sir
While running the train_pda.py I faced some issues can you please help me to debug these things.

waiting for the response...

Sir can you please upload the mer_2_30_5_0.5.h5 and prepared_data.h5 file so we can try this project.
waiting for your response...

Hi Long,

I have a small question on the loss function in the paper (i.e. Eq. 2).

I am a bit confused now and think it should be:

Moreover, in the implementation, I found two weights w0 and w1:

Are they for balancing between valid/invalid pixels?

Thx!

Thanks for your work!
May I ask which part of the datasets you download specifically? And whether you are willing to disclose your network training weights.

Thank you so much for sharing your valuable research and code.

According to Table 2 and 3 in the paper, Sparse-to-Dense [29] and MSG-CHN [25] seem to be used as the structure of 'Network 2'.
But when I look at the code train_depth.py, I can't find the parts for these two models, only 'PyramidCNN' exists.

Where can I find the code for the above two models?
Thank you.