Questions about the preprocessing code

google / dynamic-video-depth

Code for the SIGGRAPH 2021 paper "Consistent Depth of Moving Objects in Video".

https://dynamic-video-depth.github.io

Apache License 2.0

262 stars 40 forks source link

Questions about the preprocessing code #6

Open shanemankiw opened 2 years ago

shanemankiw commented 2 years ago

https://github.com/google/dynamic-video-depth/blob/7dab8f9e156fa35735301695ea020aee7221fb31/scripts/preprocess/davis/generate_frame_midas.py#L134

Is this supposed to be 'mvs_depths' instead of 'full_pred_depths'? Or the 'depth_mvs' would also be from MidasNet like depth_pred.

nnop commented 2 years ago

+1, this seems a bug.

ztzhang commented 2 years ago

Hi, this is kind of an intentional "hack" since for DAVIS, those mvs depth maps are really sparse, therefore do not visualize well in then end. Since the visualization code is written in a way where we assume a semi-dense depth maps for static regions, this makes sure things do not break when the mvs depth is from some sfm systems.