NVlabs / neuralangelo

Official implementation of "Neuralangelo: High-Fidelity Neural Surface Reconstruction" (CVPR 2023)
https://research.nvidia.com/labs/dir/neuralangelo/
Other
4.38k stars 388 forks source link

Poor Result on Custom Training Dataset #103

Open williambittner1 opened 1 year ago

williambittner1 commented 1 year ago

I have trained on a custom dataset (349 images from an indoor office scene), however the results (depth map, normal map, extracted mesh geometry) are very poor. (RGB seems quite good though). The loss and psnr graphs also seem quite suspicious. I followed your data-preprocessing instructions and the colmap-poses and points seem to be good. I also also set SCENE_TYPE=indoor during data preprocessing. Can you help me find the problem?

Screenshot 2023-09-02 150926 Screenshot 2023-09-02 150947 Screenshot 2023-09-02 150852 Screenshot 2023-09-02 151020 Screenshot 2023-09-02 150812 Screenshot 2023-08-31 132048
rafaepires commented 1 year ago

I have a same problem... williambittner1, do you find a solution? best wishes.

williambittner1 commented 1 year ago

I haven't fully solved it yet. However, I now get some better results after rerunning colmap manually and changing scene_type to object. Next, I will probably try increasing the scene_scale to 2 (cfg.data.readjust.scale = 2. inside generate_config.py) instead of 1 and run again with scene_type "indoor". Let's see...

raficabral commented 5 months ago

@williambittner1 did you have better output?