tinynerf

Concise (<1000 locs) and fast implementation of several NeRF techniques. Currently it contains an implementation of vanilla NeRF, K-Planes and Cobafa, accelerated with a single CUDA kernel to compute the weights from 'NeRF equation'.

output_video.mp4

Setup

Create a conda environment, install PyTorch, and install requirements.

conda create -n tinynerf python=3.10
conda activate tinynerf

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

pip install -r requirements.txt

I got the dataset from nerfstudio, running the ns-download-data blender command and then copying it to ./data/.

Run it:

CUDA_VISIBLE_DEVICES=0 python train.py --data data/blender/chair --method kplanes --datatype synthetic --output k_planes --batch_size 512

Features

• Vanilla NeRF, K-Planes and Cobafa
• Occupancy grid to accelerate training (based on Instant-NGP but with slightly different decaying method)
• Unbounded and AABB scenes
• Dynamic batches, each iteration process a constant number of samples by packing samples from each ray
• CUDA implementation of NeRF weights computation
• Reproduction of KPlanes results on synthetic dataset
• Reproduction of Cobafa results on synthetic dataset
• Proposal sampling
• COLMAP data loading
• Appearance embedding

References

These repositories were useful learning resources :

• https://github.com/KAIR-BAIR/nerfacc
• https://github.com/apchenstu/TensoRF
• https://github.com/sarafridov/K-Planes