NeRF

A neural radiance field (NeRF) is a fully-connected neural network that can generate novel views of complex 3D scenes, based on a partial set of 2D images. It is trained to use a rendering loss to reproduce input views of a scene. It works by taking input images representing a scene and interpolating between them to render one complete scene. NeRF is a highly effective way to generate images for synthetic data.

A NeRF network is trained to map directly from viewing direction and spatial location (5D input) to opacity and color (4D output), using volume rendering to render new views. NeRF is a computationally-intensive algorithm, and processing of complex scenes can take hours or days. However, new algorithms are available that dramatically improve performance.

Environment

pip install tqdm==4.64.0 configargparse==1.5.3 imageio==2.19.3 opencv-python==4.6.0.66

Pretrained Model

Lego:

• Coarse: 192 pts/ray: ckpt

• Coarse: 64 pts/ray + Fine: 128 pts/ray: ckpt

Fern:

• Coarse: 192 pts/ray: ckpt

• Coarse: 64 pts/ray + Fine: 128 pts/ray: ckpt

Performance

Lego	Speed	PSNR ↑	SSIM ↑	LPIPS ↓
Coarse: 192 pts/ray (torch converted ckpt)	13.37s/img	30.52	0.9565	0.032
Coarse: 192 pts/ray (Train from scratch)	13.37s/img	28.16	0.9233	0.055
Coarse: 64 pts/ray + Fine: 128 pts/ray (torch converted ckpt)	1305.28s/img	N/A	N/A	N/A
Coarse: 64 pts/ray + Fine: 128 pts/ray (Train from scratch)	1305.28s/img	N/A	N/A	N/A
Paper: 64 pts/ray + Fine: 128 pts/ray	4.69s/img	32.54	0.961	0.050

Fern	Speed	PSNR ↑	SSIM ↑	LPIPS ↓
Coarse: 192 pts/ray (torch converted ckpt)	13.37s/img	26.06	0.8305	0.095
Coarse: 192 pts/ray (Train from scratch)	13.37s/img	25.43	0.7963	0.121
Coarse: 64 pts/ray + Fine: 128 pts/ray (torch converted ckpt)	1305.28s/img	N/A	N/A	N/A
Coarse: 64 pts/ray + Fine: 128 pts/ray (Train from scratch)	1305.28s/img	N/A	N/A	N/A
Paper: 64 pts/ray + Fine: 128 pts/ray	4.69s/img	25.17	0.792	0.280

Note: the significant speed drop comes from the unsupported N-dim ms.numpy.searchsorted that instead implemented with a for loop. Related issue.

Training Parameter Description

Please check out the config script.

Example

Here, how to use NeRF model will be introduced as following.

Dataset

At first, you should download dataset by yourself. NeRF Synthetics.

After you get the dataset, make sure your path is sturctured as following:

.datasets/
└── nerf_synthetics
    └── lego
        ├── test [600 entries exceeds filelimit, not opening dir]
        ├── train [100 entries exceeds filelimit, not opening dir]
        ├── transforms_test.json
        ├── transforms_train.json
        ├── transforms_val.json
        └── val [100 entries exceeds filelimit, not opening dir]

and modify the datadir in the config file accordingly.

Train/Evaludate/Inference Model

python {train,eval}.py [data_related_options] \
[--ckpt /path/to/checkpoint] [--no_reload] [--gpu 0] \
[--render_test]
# `--no_reload`: not resume from last run.
# `--gpu`: the used gpu.
# `--ckpt`: specify the path of a checkpoint (from trained or downloaded).
# `--render_test`: render test images to perform quantitative comparison.

Fern coarse

python {train,eval}.py --name fern_coarse --data_dir ../datasets/nerf_llff/fern/ --dataset_type llff --factor 8 --llff_hold 8 --cap_n_rand 4096 --cap_n_samples 192 --cap_n_importance 0 --use_view_dirs --raw_noise_std 1e0 [--ckpt /path/to/checkpoint]

Fern full

Note: the significant speed drop comes from the unsupported N-dim ms.numpy.searchsorted that instead implemented with a for loop. Related issue.

python {train,eval}.py --name fern --data_dir ../datasets/nerf_llff/fern/ --dataset_type llff --factor 8 --llff_hold 8 --cap_n_rand 4096 --cap_n_samples 64 --cap_n_importance 128 --use_view_dirs --raw_noise_std 1e0 [--ckpt /path/to/checkpoint]

Lego coarse

python {train,eval}.py --name lego_coarse --data_dir ../datasets/nerf_synthetics/lego --dataset_type blender --half_res --cap_n_rand 4096 --cap_n_samples 192 --cap_n_importance 0 --use_view_dirs --raw_noise_std 1e0 --white_bkgd [--ckpt /path/to/checkpoint]

Lego full

Note: the significant speed drop comes from the unsupported N-dim ms.numpy.searchsorted that instead implemented with a for loop. Related issue.

python {train,eval}.py --name lego --data_dir ../datasets/nerf_synthetics/lego --dataset_type blender --half_res --cap_n_rand 4096 --cap_n_samples 64 --cap_n_importance 128 --use_view_dirs --raw_noise_std 1e0 --white_bkgd [--ckpt /path/to/checkpoint]

Results