search

yenchenlin / nerf-pytorch

A PyTorch implementation of NeRF (Neural Radiance Fields) that reproduces the results.
MIT License
5.33k stars 1.04k forks source link

The rendered img has a large area of blank ? #130

Open chenkang455 opened 9 months ago

chenkang455 commented 9 months ago

000 I adjusted the white background of the Lego to black during training, and encountered this issue in the training process. What could be the possible cause? (Here, I commented out the "white_bkgd" part.)

chenkang455 commented 9 months ago

https://github.com/yenchenlin/nerf-pytorch/assets/72788314/b271f164-a8c6-4c6e-9364-727466ba4fc4

And this is the rendered video, i wonder why the low part is perfect while the high part is almost black, could it be the reason that i comment the white regularization ? 

    # if white_bkgd:
    #     rgb_map = rgb_map + (1.-acc_map[...,None])
chenkang455 commented 9 months ago

Alternatively, there is another possibility that the issue could be due to my own Lego dataset. When generating the data, I did not follow the official data provided, but instead used Blender to create it. For example, in this image, my training data was generated by rotating the camera around the bottom, while the test data was generated by rotating the camera around the top. Could the black region in the upper half be a result of the lack of information from the camera below or the camera above? image