Preprocessing of my own depth maps

[Azpril45]

Dear @shurans

Thanks for sharing your work. I have read your paper and had some questions about the preprocessing procedure. It will be appreciated that if you can give me some help. In the paper, you said that you make some modifications to the standard TSDF and turned it to flipped TSDF which has the strong gradient on the surface, providing a more meaningful signal for the network to learn better geometric features. I want to use this flipped TSDF to generate my own volume data from a single depth image like this This is a hand pose depth image from NYU hand pose dataset.

I tried to do some modifications on your suncg_data_layer.cu layer to generate a 3d volume date. However, a parameter names camera pose confused me. I did some researches and it seems like a 4x4 matrix.

Can you tell me how can I generate a 3d volume data from my own depth image by using flipped TSDF you described or give me some hints about it?

Thank you Sincerely yours

Tony

[JamesTonG321]

Dear @Azpril45, I'm interested in your work, is that you do some modifications in suncg_data_layer.cu layer to make the standard TSDF to she flipped TSDF? Thank you.

[shurans]

Hi @Azpril45,

1. Transforming from standard tsdf (not projective tsdf) to flipped tsdf is very simple: for each voxel, if its tsdf value is (+/-)d then the flipped tsdf value is (+/-)(1-d). Therefore if you have code to compute standard tsdf you just need to change one line of code to get the flipped tsdf.
2. the camera pose is 3D transformation matrix from camera coordinate to world coordinate.

[jiunyen-ching]

Hi @Azpril45 are you able to backproject the depth map to its correct positions in the world coordinate frame?