ngp_pl

Instant-ngp (only NeRF) in pytorch+cuda trained with pytorch-lightning (high quality with high speed). This repo aims at providing a concise pytorch interface to facilitate future research, and am grateful if you can share it (and a citation is highly appreciated)!

progress.mp4

lego_trainval_5min_PNSR35.73.mp4

• Official CUDA implementation
• torch-ngp another pytorch implementation that I highly referenced.

💻 Installation

This implementation has strict requirements due to dependencies on other libraries, if you encounter installation problem due to hardware/software mismatch, I'm afraid there is no intention to support different platforms (you are welcomed to contribute).

Hardware

• OS: Ubuntu 20.04
• NVIDIA GPU with Compute Compatibility >= 75 and memory > 8GB (Tested with RTX 2080 Ti), CUDA 11.3 (might work with older version)

Software

• Clone this repo by git clone https://github.com/kwea123/ngp_pl

• Python>=3.8 (installation via anaconda is recommended, use conda create -n ngp_pl python=3.8 to create a conda environment and activate it by conda activate ngp_pl)

• Python libraries

  • Install pytorch (1.11.0) by pip install torch torchvision --extra-index-url https://download.pytorch.org/whl/cu113
  • Install tinycudann following their instruction (compilation and pytorch extension)
  • Install apex following their instruction
  • Install core requirements by pip install -r requirements.txt

• Cuda extension: Run pip install models/csrc/ --use-feature=in-tree-build

📚 Data preparation

Download preprocessed datasets from NSVF.

🔑 Training

Quickstart: python train.py --root_dir <path/to/lego> --exp_name Lego

It will train the lego scene for 20k steps (each step with 8192 rays), and perform one testing at the end. The training process should finish within about 5 minutes (saving testing image is slow, add --no_save_test to disable). Testing PSNR will be shown at the end.

More options can be found in opt.py.

Comparison with torch-ngp and the paper

I compared the quality (average testing PSNR on Synthetic-NeRF) and the inference speed (on Lego scene) v.s. the concurrent work torch-ngp (default settings) and the paper, all trained for about 5 minutes:

Method	avg PSNR	FPS
torch-ngp	31.46	7.8
mine	32.38	36.2
instant-ngp paper	33.18	60

As for quality, mine is slightly better than torch-ngp, but the result might fluctuate across different runs.

As for speed, mine is faster than torch-ngp, but is still only half fast as instant-ngp. Speed is dependent on the scene (if most of the scene is empty, speed will be faster).

^{Left: torch-ngp. Right: mine.}

More details are in the following section.

Benchmarks

To run benchmarks, use the scripts under benchmarking.

Followings are my results:

Synthetic-NeRF

	Mic	Ficus	Chair	Hotdog	Materials	Drums	Ship	Lego	AVG
PSNR	35.23	33.64	34.78	36.76	28.77	25.61	29.57	34.69	32.38
FPS	40.81	34.02	49.80	25.06	20.08	37.77	15.77	36.20	32.44

Synthetic-NSVF

	Wineholder	Steamtrain	Toad	Robot	Bike	Palace	Spaceship	Lifestyle	AVG
PSNR	31.06	35.65	34.49	36.23	36.99	36.36	35.48	33.96	35.03
FPS	47.07	75.17	50.42	64.87	66.88	28.62	35.55	22.84	48.93

Tanks and Temples

	Ignatius	Truck	Barn	Caterpillar	Family	AVG
PSNR					33.77
FPS					6.16

BlendedMVS

	Jade	Fountain	Character	Statues	AVG
PSNR			30.16	26.93
FPS			35.99	19.22

TODO

• test multi-gpu training

• support custom dataset

• benchmark on public datasets