This is the Jittor implementation of Neuralangelo: High-Fidelity Neural Surface Reconstruction.
The implementation achieves nearly identical results as the official implementation, and even slightly faster and better.
see requirements.txt.
Use the same dataset format with DTU dataset.
Just run main.py with your args.
Configs could be modified in config.py.
After training, mesh will be automatically extracted to the log directory.
Change the fast_train variable in config.py as you need.
| fast_train | GPU VRAM | Run time(RTX 4090) |
|---|---|---|
| True | 2GB | ~ 1 hour |
| False | 12GB | ~ 4 hour |
The original implementation comes from the following cool project:
Their licenses can be seen at licenses/, many thanks for their nice work!