PyTorch implementation of paper "Neural Scene Flow Fields for Space-Time View Synthesis of Dynamic Scenes", CVPR 2021
[Project Website] [Paper] [Video]
The code is tested with Python3, Pytorch >= 1.6 and CUDA >= 10.2, the dependencies includes
- configargparse
- matplotlib
- opencv
- scikit-image
- scipy
- cupy
- imageio.
- tqdm
- kornia
The current version in this github include some improvement for monocular videos in the wild. For reference code matched paper's description, please check out this branch
-
Download nerf_data.zip from link, an example input video with SfM camera poses and intrinsics estimated from COLMAP (Note you need to use COLMAP "colmap image_undistorter" command to undistort input images to get "dense" folder as shown in the example, this dense folder should include "images" and "sparse" folders).
-
Download single view depth prediction model "model.pt" from link, and put it on the folder "nsff_scripts".
-
Run the following commands to generate required inputs for training/inference:
# Usage
cd nsff_scripts
# create camera intrinsics/extrinsic format for NSFF, same as original NeRF where it uses imgs2poses.py script from the LLFF code: https://github.com/Fyusion/LLFF/blob/master/imgs2poses.py
python save_poses_nerf.py --data_path "/home/xxx/Neural-Scene-Flow-Fields/kid-running/dense/"
# Resize input images and run single view model,
# argument resize_height: resized image height for model training, width will be resized based on original aspect ratio
python run_midas.py --data_path "/home/xxx/Neural-Scene-Flow-Fields/kid-running/dense/" --resize_height 288
# Run optical flow model
./download_models.sh
python run_flows_video.py --model models/raft-things.pth --data_path /home/xxx/Neural-Scene-Flow-Fields/kid-running/dense/
- Download pretraind model "kid-running_ndc_5f_sv_of_sm_unify3_F00-30.zip" from link. Unzipping and putting it in the folder "nsff_exp/logs/kid-running_ndc_5f_sv_of_sm_unify3_F00-30/360000.tar".
Set datadir in config/config_kid-running.txt to the root directory of input video. Then go to directory "nsff_exp":
cd nsff_exp
mkdir logs
- Rendering of fixed time, viewpoint interpolation
python run_nerf.py --config configs/config_kid-running.txt --render_bt --target_idx 10
By running the example command, you should get the following result:
- Rendering of fixed viewpoint, time interpolation
python run_nerf.py --config configs/config_kid-running.txt --render_lockcam_slowmo --target_idx 8
By running the example command, you should get the following result:
- Rendering of space-time interpolation
python run_nerf.py --config configs/config_kid-running.txt --render_slowmo_bt --target_idx 10
By running the example command, you should get the following result:
- In configs/config_kid-running.txt, modifying expname to any name you like (different from the original one), and running the following command to train the model:
python run_nerf.py --config configs/config_kid-running.txt
The per-scene training takes ~2 days using 4 Nvidia GTX2080TI GPUs.
- Several parameters in config files you might need to know for training a good model on in-the-wild video
- final_height: this must be same as --resize_height argument in run_midas.py, in kid-running case, it should be 288.
- N_samples: in order to render images with higher resolution, you have to increase number sampled points such as 256 or 512
- chain_sf: model will perform local 5 frame consistency if set True, and perform 3 frame consistency if set False. For faster training, setting to False.
- start_frame, end_frame: indicate training frame range. The default model usually works for video of 1~2s and 30-60 frames work the best for default hyperparameters. Training on longer frames can cause oversmooth rendering. To mitigate the effect, you can increase the capacity of the network by increasing netwidth to 512.
- decay_iteration: number of iteartion in initialization stage. Data-driven losses will decay every 1000 * decay_iteration steps. We have updated code to automatically calculate number of decay iterations.
- no_ndc: our current implementation only supports reconstruction in NDC space, meaning it only works for forward-facing scene, same as original NeRF.
- use_motion_mask, num_extra_sample: whether to use estimated coarse motion segmentation mask to perform hard-mining sampling during initialization stage, and how many extra samples during initialization stage.
- w_depth, w_optical_flow: weight of losses for single-view depth and geometry consistency priors described in the paper. Weights of (0.4, 0.2) or (0.2, 0.1) usually work the best for most of the videos.
- If you see signifacnt ghosting result in the final rendering, you might try the suggestion from link
-
Download Dynamic Scene dataset "dynamic_scene_data_full.zip" from link
-
Download pretrained model "dynamic_scene_pretrained_models.zip" from link, unzip and put them in the folder "nsff_exp/logs/"
-
Run the following command for each scene to get quantitative results reported in the paper:
# Usage: configs/config_xxx.txt indicates each scene name such as config_balloon1-2.txt in nsff/configs
python evaluation.py --config configs/config_xxx.txt
- Note: you have to use modified LPIPS implementation included in this branch in order to measure LIPIS error for dynamic region only as described in the paper.
The code is based on implementation of several prior work:
- https://github.com/sniklaus/softmax-splatting
- https://github.com/yenchenlin/nerf-pytorch
- https://github.com/JKOK005/dVRK-Linear-Interpolator-
- https://github.com/richzhang/PerceptualSimilarity
- https://github.com/intel-isl/MiDaS
- https://github.com/princeton-vl/RAFT
- https://github.com/NVIDIA/flownet2-pytorch
This repository is released under the MIT license.
If you find our code/models useful, please consider citing our paper:
@InProceedings{li2020neural,
title={Neural Scene Flow Fields for Space-Time View Synthesis of Dynamic Scenes},
author={Li, Zhengqi and Niklaus, Simon and Snavely, Noah and Wang, Oliver},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
year={2021}
}