NeRF++

Codebase for arXiv preprint:

• Work with 360 capture of large-scale unbounded scenes.
• Support multi-gpu training and inference with PyTorch DistributedDataParallel (DDP).
• Optimize per-image autoexposure (experimental feature).

Data

• Download our preprocessed data from tanks_and_temples, lf_data.
• Put the data in the sub-folder data/ of this code directory.
• Data format.
  • Each scene consists of 3 splits: train/test/validation.
  • Intrinsics and poses are stored as flattened 4x4 matrices (row-major).
  • Poses are camera-to-world, not world-to-camera transformations.
  • Opencv camera coordinate system is adopted, i.e., x--->right, y--->down, z--->scene.
  • To convert camera poses between Opencv and Opengl conventions, the following snippet can be used for both Opengl2Opencv and Opencv2Opengl.

    import numpy as np
    def convert_pose(C2W):
        flip_yz = np.eye(4)
        flip_yz[1, 1] = -1
        flip_yz[2, 2] = -1
        C2W = np.matmul(C2W, flip_yz)
        return C2W
    

  • Scene normalization: move the average camera center to origin, and put all the camera centers inside the unit sphere.

Create environment

conda env create --file environment.yml
conda activate nerf-ddp

Training (Use all available GPUs by default)

python ddp_train_nerf.py --config configs/tanks_and_temples/tat_training_truck.txt

Testing (Use all available GPUs by default)

python ddp_test_nerf.py --config configs/tanks_and_temples/tat_training_truck.txt \
                        --render_splits test,camera_path