This is the official repository for the project "Non-Rigid Neural Radiance Fields: Reconstruction and Novel View Synthesis of a Dynamic Scene From Monocular Video" (NR-NeRF). We extend NeRF, a state-of-the-art method for photorealistic appearance and geometry reconstruction of a static scene, to dynamic/deforming/non-rigid scenes. For details, we refer to the preprint and the project page, which also includes supplemental videos.

We modified the losses and added multi-view support, some simple scene-editing capabilities, evaluation code, and a naive baseline in a February 2021 update. For the original version of this code, please refer to the december2020 branch of this repo.

• Clone this repository.
• (Optional) Install Miniconda.
• Setup the conda environment nrnerf (or install the requirements using pip):

conda env create -f environment.yml

• (Optional) For data loading and camera parameter estimation, we have included a dummy implementation that only works on the included example sequence. If you do not want to write your own implementation as specified at the end of this README, you can instead use the following programs and files:
  • Install COLMAP.
  • From nerf-pytorch, use load_llff.py to replace the example version included in this repo.
    • In load_llff_data(), replace sc = 1. if bd_factor is None else 1./(bds.min() * bd_factor) with sc = 1./(bds.max() - bds.min())
  • From LLFF, copy from llff/poses/ the three files colmap_read_model.py, colmap_wrapper.py, and pose_utils.py directly into ./llff_preprocessing (replacing existing files).
    • In pose_utils.py fix the imports by:
      • Commenting out import skimage.transform,
      • Replacing from llff.poses.colmap_wrapper import run_colmap with from .colmap_wrapper import run_colmap,
      • Replacing import llff.poses.colmap_read_model as read_model with from . import colmap_read_model as read_model.
• (Optional) An installation of FFMPEG enables automatic video generation from images and frame extraction from video input.

conda install -c conda-forge ffmpeg

Having set up the environment, we now show an example that starts with a folder of just images and ends up with a fixed viewpoint re-rendering of the sequence. Please read the sections after this one for details on each step and how to adapt the pipeline to other sequences.

We first navigate into the parent folder (where train.py etc. lie) and activate the conda environment:

conda activate nrnerf

(Preprocess) We then determine the camera parameters:

python preprocess.py --input data/example_sequence/

(Training) Next, we train the model with the scene-specific config:

python train.py --config configs/example_sequence.txt

(Free Viewpoint Rendering) Finally, we synthesize a novel camera path:

python free_viewpoint_rendering.py --input experiments/experiment_1/ --deformations train --camera_path fixed --fixed_view 10

All results will be in the same folder, experiments/experiment_1/output/train_fixed_10/.

Overall, the input video (left) is re-rendered into a fixed novel view (right):

• Works with video file input,
• Script for lens distortion estimation and undistortion of input files,
• Automatic multi-GPU support (torch.nn.DataParallel),
• Automatically continues training if previous training detected,
• Some modifications to lessen GPU memory requirements and to speed-up loading at the start of training.

As this is a research project, it is not sufficiently robust to work on arbitrary scenes. Here are some tips to consider when recordings new scenes:

• Sequences should have lengths of about 100-300 frames. More frames require longer training.
• Avoid blur (e.g., motion blur or out-of-focus blur).
• Keep camera settings like color temperature and focal length fixed.
• Avoid lens distortions or estimate distortion parameters for undistortion.
• Stick to front-facing camera paths that capture most of the scene in all images.
• Use sufficient lighting and avoid changing it while recording.
• Avoid self-shadowing.
• Only record Lambertian surfaces, avoid view-dependent effects like specularities (view-dependent effects can be activated by setting use_viewdirs=True).
• The background needs to be static and dominant enough for SfM to estimate extrinsics.
• Limited scene size: Ensure that the background is not more than an order of magnitude further from the camera compared to the non-rigid foreground.

Determining Camera Parameters

Before we can train a network on a newly recorded sequence, we need to estimate its camera parameters (extrinsics and intrinsics).

The preprocessing code assumes the folder structure PARENT_FOLDER/images/IMAGE_NAME1.png. To determine the camera parameters for such a sequence, please run

python preprocess.py --input PARENT_FOLDER

The --output OUTPUT_FOLDER option allows to set a custom output folder, otherwise PARENT_FOLDER is used by default.

(Optional) Lens Distortion Estimation and Image Undistortion

While not necessary for decent results with most camera lenses, the preprocessing code allows to estimate lens distortions from a checkerboard/chessboard sequence and to then use the estimated distortion parameters to undistort input sequences recorded with the same camera.

First, record a checkerboard sequence and run the following command to estimate lens distortion parameters from it:

python preprocess.py --calibrate_lens_distortion --input PARENT_FOLDER --checkerboard_width WIDTH --checkerboard_height HEIGHT

The calibration code uses OpenCV. HEIGHT and WIDTH refer to the number of squares, not to lengths. The optional flags --visualize_detections and --undistort_calibration_images might help with determining issues with the calibration process, see preprocess.py for details.

Then, in order to undistort an input sequence using the computed parameters, simply add --undistort_with_calibration_file PATH_TO_LENS_DISTORTION_JSON when preprocessing the sequence using preprocess.py as described under Determining Camera Parameters.

(Optional) Video Input

In addition to image files, the preprocessing code in preprocess.py also supports video input. Simply set --input to the video file.

This requires an installation of ffmpeg. The --ffmpeg_path PATH_TO_EXECUTABLE option allows to set a custom path to an ffmpeg executable.

The --fps 10 option can be used to modify the framerate at which images are extracted from the video. The default is 5.

The config file default.txt needs to be modified as follows:

• rootdir: An output folder that collects all experiments (i.e. multiple trainings)
• datadir: Recorded input sequence. Set to PARENT_FOLDER from the Preprocess section above
• expname: Name of this experiment. Output will be written to rootdir/expname/

Other relevant parameters are:

• offsets_loss_weight, divergence_loss_weight, rigidity_loss_weight: Weights for loss terms. Need to be tuned for each scene, see the preprint for details.
• factor: Downsamples the input sequence by factor before training on it.
• use_viewdirs: Set to True to activate view-dependent effects. Note that this slows down training by about 20% (approximate) or 35% (exact) on a V100 GPU.
• approx_nonrigid_viewdirs: True uses a fast finite difference approximation of the view direction, False computes the exact direction.

Finally, start the training by running:

python train.py

A custom config file can optionally be passed via --config CONFIG_FILE.

The train_block_size and test_block_size options allow to split the images into training and test blocks. The scheme is AAAAABBAAAAABBAAA for train_block_size=5 and test_block_size=2. Note that optimizing for the latent codes of test images slows down training by about 30% (relative to only using training images) due to an additional backwards pass.

If a previous version of the experiment exists, train.py will automatically continue training from it. To prevent that, pass the --no_reload flag.

The --time_conditioned_baseline flag will turn on the naive NR-NeRF baseline described in the paper.

(Optional) Multi-View Data and Per-Camera Intrinsics

NR-NeRF supports multi-view setups. A deformation latent code is associated with each time step, and multiple images can be associated with each time step.

To know which images belong to the same time step, we use image_to_camera_id_and_timestep.json in the data PARENT_FOLDER.

You can create the JSON file by first filling a image_to_camera_id_and_timestep dict in Python:

image_to_camera_id_and_timestep[image_filename] = (camera_string, timestep_int)

where image_filename is one of the filenames in PARENT_FOLDER/images/, camera_string is an arbitrary string that identifies a camera, and timestep_int is an integer that identifies the time step. Then write the dict to a JSON file as follows:

import json
with open(PARENT_FOLDER + "image_to_camera_id_and_timestep.json", "w") as json_file:
    json.dump(image_to_camera_id_and_timestep, json_file, indent=4)

Example file content with one camera and two time steps:

{
    "image00000000.jpg": [
        "0",
        8
    ],
    "image00000001.jpg": [
        "0",
        9
    ]
}

To load multi-view data, please replace the call to load_llff_data in train.py with a call to load_llff_data_multi_view in the provided load_llff.py. Note that downsampling images by a factor is not implemented in load_llff_data_multi_view.

Furthermore, using multi-view data requires specifying per-camera intrinsics. In the implementation, the file PARENT_FOLDER/calibration.json contains both intrinsics and extrinsics. Note that this calibration.json file is not necessary for monocular setups. If you want to use per-image intrinsics for a monocular setup, treat it as a multi-view setup where each image has its own unique time step.

Similar to before, calibration.json starts out as a Python dict calibration. It contains entries calibration["min_bound"] = min_bound and calibration["max_bound"] = max_bound that each are a float. min_bound is the distance from a camera to the near plance, and max_bound is the distance to the far plane, both in world units. In addition, for each camera, calibration contains an entry calibration[camera_string] = camera_calibration, where camera_string has to be consistent with the camera_string from image_to_camera_id_and_timestep. camera_calibration is again a Python dict that contains an entry translation (list of three floats), rotation (list of three lists of three floats, gotten via rotation.tolist() from a numpy array), center_x and center_y (each a float, camera center in pixel units), focal_x and focal_y (each a float, focal length in pixel units), and height and width (each an integer, in pixel units). For translation and rotation, see at the end of this README for how they are defined.

Example file content with one camera named 0:

{
    "min_bound": 0.0,
    "max_bound": 2.0189487179595886,
    "0": {
        "translation": [
            -0.041070333333333334,
            1.1753333333333333,
            0.49935666666666667
        ],
        "rotation": [
            [
                0.0577962,
                -0.997661,
                -0.0364925
            ],
            [
                0.558001,
                0.00197212,
                0.829838
            ],
            [
                -0.827825,
                -0.0683243,
                0.55681
            ]
        ],
        "center_x": 2572.48,
        "center_y": 1875.78,
        "focal_x": 5363.46,
        "focal_y": 5363.46,
        "height": 3840,
        "width": 5120
    }
}

Once we've trained a network, we can render it into novel views.

The following arguments are mandatory:

• input: Set to the folder of the trained network, i.e. rootdir/expname/
• deformations: Set to the subset of the deformations/images that are to be used. Can be train, test, or all
• camera_path: Possible camera paths are: input_recontruction, fixed, and spiral.

Then, we can synthesize novel views by running:

python free_viewpoint_rendering.py --input INPUT --deformations train --camera_path fixed

The fixed camera view uses the first input view by default. This can be set to another index (e.g. 5) with --fixed_view 5.

Evaluation

When using fixed, the code will compute the standard deviation across images to visualize the background stability and store the result in standard_deviation.png.

When using input_recontruction, the code will automatically evaluate the resulting renderings quantitatively against the groundtruth and store the result in scores.json. In order to optionally use LPIPS during evaluation, you can clone the required code (in the main folder):

git clone https://github.com/richzhang/PerceptualSimilarity

Simple Scene Editing

The --motion_factor 1.2 option exaggerates the deformation by a factor of 1.2. Factors below 1 dampen the motion.

The --foreground_removal 0.01 option removes points that are more non-rigid than the provided threshold.

Furthermore, the forced background stabilization described in the preprint can be used by passing a threshold via the --forced_background_stabilization 0.01 option.

free_viewpoint_rendering.py contains code for time interpolation that can be easily adapted to a specific scenario. Search for example with time interpolation from a fixed camera view in the file.

The canonical model (without any ray bending applied) can be rendered by setting the --render_canonical flag.

(Optional) Video Generation

For automatic video generation, please install ffmpeg. The framerate of the generated output videos can be set with --output_video_fps 5.

(Optional) Adaptive Spiral Camera Path

It is possible to use a spiral camera path that adapts to the length of the video. If you do not want to implement such a path yourself, you can copy and modify the else branch in load_llff_data of LLFF's load_llff.py. You can find a recommended wrapper in free_viewpoint_rendering: _spiral_poses. Set N_views to num_poses. We recommend multiplying rads in render_path_spiral right before the for loop by 0.5.

When using this code, please cite our preprint Tretschk et al.: Non-Rigid Neural Radiance Fields as well as the following works on which it builds:

@misc{tretschk2020nonrigid,
      title={Non-Rigid Neural Radiance Fields: Reconstruction and Novel View Synthesis of a Dynamic Scene From Monocular Video},
      author={Edgar Tretschk and Ayush Tewari and Vladislav Golyanik and Michael Zollhöfer and Christoph Lassner and Christian Theobalt},
      year={2020},
      eprint={2012.12247},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}
@misc{lin2020nerfpytorch,
  title={NeRF-pytorch},
  author={Yen-Chen, Lin},
  howpublished={\url{https://github.com/yenchenlin/nerf-pytorch/}},
  year={2020}
}
@inproceedings{mildenhall2020nerf,
 title={NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis},
 author={Ben Mildenhall and Pratul P. Srinivasan and Matthew Tancik and Jonathan T. Barron and Ravi Ramamoorthi and Ren Ng},
 year={2020},
 booktitle={ECCV},
}

load_llff_data from load_llff.py needs to return a numpy array images of shape N x H x W x 3 with RGB values scaled to lie between 0 and 1, a numpy array poses of shape N x 3 x 5, where poses[:,:,:3] are the camera extrinsic rotations, poses[:,:,3] are the camera extrinsic translations in world units, and poses[:,:,4] are height, width, focal length in pixels at every frame (the same at all N frames), bds is a numpy array containing the depth values of near and far planes in world units (only the maximum and minimum entries of bds matter), render_poses is a numpy array of shape N x 3 x 4 with rotation and translation encoded as for poses, and i_test is an image index. The first argument specifies the directory from which the images should be loaded, and the second argument specifies a downsampling factor that should be applied to the images. The remaining arguments can be ignored.

gen_poses from llff_preprocessing/pose_utils.py should compute and store camera parameters of the images given by the first argument such that the format is compatible with load_llff_data. The second argument can be ignored.

The camera extrinsic translation is in world space. The translations should be scaled such that the overall scene roughly lies in the unit cube. The camera extrinsic rotation is camera-to-world, R * c = w. The camera coordinate system has the x-axis pointing to the right, y up, and z back.

This code builds on the PyTorch port by Yen-Chen Lin of the original NeRF code. Both are released under an MIT license. Several functions in run_nerf_helpers.py are modified versions from the FFJORD code, which is released under an MIT license. We thank all of them for releasing their code.

We release this code under an MIT license as well. You can find all licenses in the file LICENSE.