SinNeRF: Training Neural Radiance Fields on Complex Scenes from a Single Image

[Paper] [Website]

Pipeline

Code

Environment

pip install -r requirements.txt

Dataset Preparation

Please download the datasets from these links:

• NeRF synthetic: Download nerf_synthetic.zip from https://drive.google.com/drive/folders/128yBriW1IG_3NJ5Rp7APSTZsJqdJdfc1
• LLFF: Download nerf_llff_data.zip from https://drive.google.com/drive/folders/128yBriW1IG_3NJ5Rp7APSTZsJqdJdfc1
• DTU: Download the preprocessed DTU training data from https://drive.google.com/file/d/1eDjh-_bxKKnEuz5h-HXS7EDJn59clx6V/view

Please download the depth from here: https://drive.google.com/drive/folders/13Lc79Ox0k9Ih2o0Y9e_g_ky41Nx40eJw?usp=sharing

Training

If you meet OOM issue, try:

1. enable precision=16
2. reduce the patch size --patch_size (or --patch_size_x, --patch_size_y) and enlarge the stride size --sH, --sW

NeRF synthetic

• Step 1

  python train.py  --dataset_name blender_ray_patch_1image_rot3d  --root_dir  ../../dataset/nerf_synthetic/lego   --N_importance 64 --img_wh 400 400 --num_epochs 2000 --batch_size 1  --optimizer adam --lr 2e-4  --lr_scheduler steplr --decay_step 500 1000 --decay_gamma 0.5  --exp_name lego_s6 --with_ref --patch_size 64 --sW 6 --sH 6 --proj_weight 1 --depth_smooth_weight 0  --dis_weight 0 --num_gpus 4 --load_depth --depth_type nerf --model sinnerf --depth_weight 8 --vit_weight 10 --scan 4
  

• Step 2

  python train.py  --dataset_name blender_ray_patch_1image_rot3d  --root_dir  ../../dataset/nerf_synthetic/lego   --N_importance 64 --img_wh 400 400 --num_epochs 2000 --batch_size 1  --optimizer adam --lr 5e-5  --lr_scheduler steplr --decay_step 500 1000 --decay_gamma 0.5  --exp_name lego_s6_4ft --with_ref --patch_size 64 --sW 4 --sH 4 --proj_weight 1 --depth_smooth_weight 0  --dis_weight 0.01 --num_gpus 4 --load_depth --depth_type nerf --model sinnerf --depth_weight 8 --vit_weight 0 --pt_model xxx.ckpt --nerf_only  --scan 4
  

LLFF

• Step 1

  python train.py  --dataset_name llff_ray_patch_1image_proj  --root_dir  ../../dataset/nerf_llff_data/room   --N_importance 64 --img_wh 504 378 --num_epochs 2000 --batch_size 1  --optimizer adam --lr 2e-4  --lr_scheduler steplr --decay_step 500 1000 --decay_gamma 0.5  --exp_name llff_room_s4 --with_ref --patch_size_x 63 --patch_size_y 84 --sW 4 --sH 4 --proj_weight 1 --depth_smooth_weight 0  --dis_weight 0 --num_gpus 4 --load_depth --depth_type nerf --model sinnerf --depth_weight 8 --vit_weight 10
  

• Step 2

  python train.py  --dataset_name llff_ray_patch_1image_proj  --root_dir  ../../dataset/nerf_llff_data/room   --N_importance 64 --img_wh 504 378 --num_epochs 2000 --batch_size 1  --optimizer adam --lr 5e-5  --lr_scheduler steplr --decay_step 500 1000 --decay_gamma 0.5  --exp_name llff_room_s4_2ft --with_ref --patch_size_x 63 --patch_size_y 84 --sW 2 --sH 2 --proj_weight 1 --depth_smooth_weight 0  --dis_weight 0.01 --num_gpus 4 --load_depth --depth_type nerf --model sinnerf --depth_weight 8 --vit_weight 0 --pt_model xxx.ckpt --nerf_only
  

DTU

• Step 1

  python train.py  --dataset_name dtu_proj  --root_dir  ../../dataset/mvs_training/dtu   --N_importance 64 --img_wh 640 512 --num_epochs 2000 --batch_size 1  --optimizer adam --lr 2e-4  --lr_scheduler steplr --decay_step 500 1000 --decay_gamma 0.5  --exp_name dtu_scan4_s8 --with_ref --patch_size_y 70 --patch_size_x 56 --sW 8 --sH 8 --proj_weight 1 --depth_smooth_weight 0  --dis_weight 0 --num_gpus 4 --load_depth --depth_type nerf --model sinnerf --depth_weight 8 --vit_weight 10 --scan 4
  

• Step 2

  python train.py  --dataset_name dtu_proj  --root_dir  ../../dataset/mvs_training/dtu   --N_importance 64 --img_wh 640 512 --num_epochs 2000 --batch_size 1  --optimizer adam --lr 5e-5  --lr_scheduler steplr --decay_step 500 1000 --decay_gamma 0.5  --exp_name dtu_scan4_s8_4ft --with_ref --patch_size_y 70 --patch_size_x 56 --sW 4 --sH 4 --proj_weight 1 --depth_smooth_weight 0  --dis_weight 0.01 --num_gpus 4 --load_depth --depth_type nerf --model sinnerf --depth_weight 8 --vit_weight 0 --pt_model xxx.ckpt --nerf_only  --scan 4
  

More finetuning with smaller strides benefits reconstruction quality.

Testing

python eval.py  --dataset_name llff  --root_dir /dataset/nerf_llff_data/room --N_importance 64 --img_wh 504 378 --model nerf --ckpt_path ckpts/room.ckpt --timestamp test

Please use --split val for NeRF synthetic dataset.

Acknowledgement

Codebase based on https://github.com/kwea123/nerf_pl . Thanks for sharing!

Citation

If you find this repo is helpful, please cite:

@InProceedings{Xu_2022_SinNeRF,
author = {Xu, Dejia and Jiang, Yifan and Wang, Peihao and Fan, Zhiwen and Shi, Humphrey and Wang, Zhangyang},
title = {SinNeRF: Training Neural Radiance Fields on Complex Scenes from a Single Image},
journal={arXiv preprint arXiv:2204.00928},
year={2022}
}