[Unofficial code-base] NeRF--: Neural Radiance Fields Without Known Camera Parameters

[ Project | Paper | Official code base ] ⬅️ Thanks the original authors for the great work!

• ⚠️ This is an ❗ unofficial ❗ pytorch re-implementation of the paper NeRF--: Neural Radiance Fields Without Known Camera Parameters.
• I have reproduced the results on the LLFF-fern dataset, LLFF-flower dataset, personal photos, and some YouTube video clips chosen by myself.
• This repo contains implementation of both the original paper and my personal modifications & refinements.

Example results

Source 1: random YouTube video clips, time from 00:10:36 to 00:10:42

	ReLU-based NeRF-- (no refinement, stuck in local minima)	SIREN-based NeRF-- (no refinement)
input 32 raw photos, sampled at 5 fps 32 x [540 x 960 x 3]
learned scene model size	1.7 MiB / 158.7k params 8+ MLPs with width of 128
learned camera poses
predicted rgb (appearance) (with novel view synthesis)
predicted depth (geometry) (with novel view synthesis)

Source 2: random YouTube video clips, time from 00:46:17 to 00:46:28

	ReLU-based NeRF-- (with refinement, still stuck in local minima)	SIREN-based NeRF-- (with refinement)
input 27 raw photos, sampled at 2.5 fps 27 x [540 x 960 x 3]
learned scene model size	1.7 MiB / 158.7k params 8+ MLPs with width of 128
learned camera poses
predicted rgb (appearance) (with novel view synthesis)
predicted depth (geometry) (with novel view synthesis)

Source 3: photos by @crazyang

	ReLU-based NeRF-- (no refinement)	SIREN-based NeRF-- (no refinement)
input 22 raw photos 22 x [756 x 1008 x3]
learned scene model size	1.7 MiB / 158.7k params 8+ MLPs with width of 128
learned camera poses
predicted rgb (appearance) (with novel view synthesis)
predicted depth (geometry) (with novel view synthesis)

Notice that the reflectance of the piano's side is misunderstood as transmittance, which is reasonable and acceptable since no prior of the shape of the piano is provided.

What is NeRF and what is NeRF--

NeRF

NeRF is a neural (differentiable) rendering framework with great potentials. Please view [NeRF Project Page] for more details.

It represents scenes as a continuous function (typically modeled by several layers of MLP with non-linear activations); the same ideas within DeepSDF, SRN, DVR, and so on.

It is suggested to refer to [awesome-NeRF] and [awesome-neural-rendering] to catch up with recent 'exploding' development in these areas.

NeRF--

NeRF-- modifies the original NeRF from requiring known camera parameters to supporting unknown and learnable camera parameters.

• NeRF-- does the following work in the training process:

  • Joint optimization of
    • camera intrinsics
    • camera extrinsics
    • a NeRF model (appearance and geometry)
  • Using pure raw real-world images and using just photometric loss (image reconstruction loss).

• SfM+MVS

  • In other words, NeRF-- tackles exactly the same problem with what a basic SfM+MVS system like COLMAP does, but learns the camera parameters, geometry and appearance of the scene simultaneously in a more natural and holistic way, requiring no hand-crafted feature extraction procedures like SIFT or points, lines, surfaces etc.

• How?

  • Since NeRF is a neural rendering framework (which means the whole framework is differentiable), one can directly compute the gradients of the photometric loss with respect to the camera parameters.

• 🚀 Wide future of NeRF-based framework --- vision by inverse computer graphics

  • Expect more to come! Imagine direct computing of gradients of photometric loss w.r.t. illumination? object poses? object motion? object deformation? objects & background decomposition ? object relationships?...

My modifications & refinements / optional features

This repo first implements NeRF-- with nothing changed from the original paper. But it also support the following optional modifications, and will keep updating.

All the options are configured using yaml configuration files in the configs folder. See details about how to use these configs in the configuration section.

SIREN-based NeRF as backbone

Replace the ReLU activations of NeRF with sinusoidal(sin) activation. Codes borrowed and modified from [lucidrains' implementation of pi-GAN]. Please refer to SIREN and pi-GAN for more theoretical details.

To config:

model:
  framework: SirenNeRF # options: [NeRF, SirenNeRF]

📌 SIREN-based NeRF compared with ReLU-based NeRF

• SirenNeRF could lead to smoother learned scenes (especially smoother shapes). (below, left for ReLU-based, right for SIREN-based)

ReLU-based NeRF-- (no refinement)	SIREN-based NeRF-- (no refinement)

• SirenNeRF could lead to better results (smaller losses at convergence, better SSIM/PSNR metrics) .

The above two conclusions are also evidenced by the DeepSDF results shown in the SIREN project.

• SirenNeRF is a little bad for scenes with lots of sharp and messy edges (for its continuously differentiable & smoothing nature).

e.g. LLFF-flower scene

ReLU-based NeRF-- (with refinement)	SIREN-based NeRF-- (with refinement)

Note: since the raw output of SirenNeRF is relatively slower to grow, I multiply the raw output (sigma) of SirenNeRF with a factor of 10 30. To config, use model:siren_sigma_mul

[WIP] Perceptual model

For fewer shots with large viewport changes, I add an option to use a perceptual model (CLIP) and an additional perceptual loss along with the reconstruction loss, as in DietNeRF.

To config:

data:
  N_rays: -1 # options: -1 for whole image and no sampling, a integer > 0 for number of ray samples
training:
  w_perceptual: 0.01 # options: 0. for no perceptual model & loss, >0 to enable

Note: as the CLIP model requires at least 224x224 resolution and a whole image (not sampled rays) as input

• data:N_rays must set to -1 for generating whole images when training
• data:downscale must set to a proper value, and a GPU with larger memory size is required
  • or proper up-sampling is required

More choices of rotation representations / intrinsics parameterization

• rotation representation

Refer to this paper for theoretical suggestions for different choices of SO(3) representations.

To config:

model:
  so3_representation: 'axis-angle' # options: [quaternion, axis-angle, rotation6D]

• intrinsics parameterization

To config:

model:
  intrinsics_representation: 'square' # options: [square, ratio, exp]

Usage

hardware

• 💻 OS: tested on Ubuntu 16 & 18
• GPU (all assume 1024 ray samples(by default) + 756 x 1000 resolution + 128 network width)
  • scene model parameter size
    • 🌟 1.7 MiB for float32.
    • For NeRF scene model, it's just 8+ layers of MLPs with ReLU/sin activation, with width of 128.
  • 🕐 training time on 2080Ti
    • <10 mins or 0-200 epochs: learning poses mainly, and rough appearances
    • ~4 hours, from ~300 to 10000 epochs: the poses has little further changes; the NeRF model learns fine details (geometry & appearance) of the scene
  • GPU memory:
    • (training) ~3300 MiB GPU memory usage
    • (testing / rendering) lower GPU memory usage, but potentially more GPU usage since testing is on full resolution, while training is on a small batch of sampled pixel for each iteration.

software

• Python >= 3.5

• To install requirements, run:

  • Simply just run: (suggested used in anaconda environments):

    ## install torch & cuda & torchvision using your favorite tools, conda/pip
    # pip install torch torchvision
    
    ## install other requirements
    pip install numpy pyyaml addict imageio imageio-ffmpeg scikit-image tqdm tensorboardX "pytorch3d>=0.3.0" opencv-python

  • Or

    conda env create -f environment.yml

• Before running any python scripts for the first time, cd to the project root directory and add the root project directory to the PYTHONPATH by running:

  cd /path/to/improved-nerfmm
  source set_env.sh

configuration

There are three choices for giving configuration values:

• [DO NOT change] configs/base.yaml contains all the default values of the whole config.
• [Your playground] Specific config yamls in configs folder are for specific tasks. You only need to put related config keys here. It is given to the python scripts using python xxx.py --config /path/to/xxx.yaml.
  • e.g. configs/fern.yaml for training LLFF-fern datasets.
• You can also give additional runtime command-line arguments with python xxx.py --xxx:yyy val, to change the config dict: 'xxx':{'yyy': val}

The configuration overwriting priority order:

• command line args >>overwrites>> --config /path/to/xxx.yaml >>overwrites>> configs/base.yaml

data

pre-trained models

You can get pre-trained models in either of the following two ways:

• Clone the repo using git-lfs, and you will get the pre-trained models and configs in pretrained folder.
• From pretrained folder in the google-drive

Training

Before running any python scripts for the first time, cd to the project root directory and add the root project directory to the PYTHONPATH by running:

cd /path/to/improved-nerfmm
source set_env.sh

Train on example data (without refinement)

Download LLFF example data using the scripts (run in project root directory)

bash dataio/download_example_data.sh

Start training:

python train.py --config configs/fern.yaml

• To specify used GPUs: (e.g. 2 and 3) (in most cases, one GPU is quite enough.)

  python train.py --config configs/fern.yaml --device_ids 2,3

  • Currently, this repo use torch.DataParallel for multi-GPU training.

• View the training logs and stats output in the experiment folder: ./logs/fern

• Run tensorboard to monitor the training process:

  tensorboard --logdir logs/fern/events

• To resume previously interrupted training:

  python train.py --load_dir logs/fern

  • Note: Full config is automatically backed up in the experiment directory when start training. Thus when loading from a directory, the scripts will only read from your_exp_dir/config.yaml, and configs/base.yaml will not be used.

🚀 Train on your own data

• 📌 Note on suitable input

  • ①static scene ②forward-facing view ③with small view-port changes.
    • Smaller viewport change / forward facing views
      • So that a certain face of a certain object should appear in all views
      • Otherwise the training would fail in the early stages (failed to learn reasonable camera poses, and hence no chance for the NeRF).
      • This is mostly because it processes all input images at once.
    • No moving / deforming objects. (e.g. a car driving across the street, a bird flying across the sky, people waving hands)
    • No significant illumination/exposure changes. (e.g. camera moving from pointing towards the sun to back to the sun)
    • No focal length changes. Currently assume all input share the same camera intrinsics.
    • Just temporarily! (All imaginable limitations have imaginable solutions. Stay tuned!)

• 📌 Note on training

  • When training with no refinement, the training process is roughly split into two phases:
    • [0 to about 100-300 epochs] The NeRF model learns some rough blurry pixel blocks, and these rough blocks helps with optimizing the camera extrinsics.
    • [300 epochs+ to end] The camera extrinsics are almost fixed, with very small further changes; the NeRF model learns the fine details of the scene.
  • You should monitor the early 100~300 epochs of the training process. If no meaningful camera poses (especially the camera translation on xy-plane) are learned during this early stages, there almost won't be any miracle further.
  • I have not tested on >50 images, but you can give it a try.

• First, prepare your photos and put them into one separate folder, say /path/to/your_photos/xxx.png.

• Second:

  • Write a new config file for your data: (you can put any config key mentioned in configs/base.yaml)

    expname: your_expname
    data:
      data_dir: /path/to/your_photos
      #=========
      N_rays: 1024 # numer of sampled rays in training.
      downscale: 4.
      #=========

  • And run

    python train.py --config /path/to/your_config.yaml

  • Or you can use some existing config file and run:

    python train.py --config /path/to/xxx.yaml --data:data_dir /path/to/your_photos --expname your_expname

• The logging and stats would be in logs/your_expname folder.

• Monitor the training process with:

  tensorboard --logdir logs/your_expname/events

Train on video clips

• First, clip your video.mp4 with ffmepg.

  ffmpeg -ss 00:10:00 -i video.mp4 -to 00:00:05 -c copy video_clip.mp4

  • Note:

    • time format: hh:mm:ss.xxx
    • -ss means starting timestamp
    • -to means duration length, not end timestamp.

• Second, convert video_clip.mp4 into images:

  mkdir output_dir
  ffmpeg -i video_clip.mp4 -filter:v fps=fps=3/1 output_dir/img-%04d.png

  • Note:

    • 3/1 means 3 frames per second. 3 is the nominator, 1 is the denominator.

• Then train on your images with instructions in 🚀 ​Train on your own data

Automatic training with a pre-train stage and refine stage

Run

python train.py --config ./configs/fern_prefine.yaml

Or

python train.py --config ./configs/fern.yaml --training:num_epoch_pre 1000 --expname fern_prefine

You can also try on your own photos using similar configurations.

Refining a pre-trained NeRF--

This is the step suggested by original NeRF-- paper: drop all pre-trained parameters except for camera parameters, and refine.

For example, refine a pre-trained LLFF-fern scene, with original config stored in ./configs/fern.yaml, a pre-trained checkpoint in ./logs/fern/ckpts/final_xxxx.pt, and with a new experiment name fern_refine:

python train.py --config ./configs/fern.yaml --expname fern_refine --training:ckpt_file ./logs/fern/ckpts/final_xxxx.pt  --training:ckpt_only_use_keys cam_params

Note:

• --training:ckpt_only_use_keys cam_params is used to drop all the keys in the pre-trained state_dict except cam_params when loading the checkpoints.
  • Some warnings like Could not find xxx in checkpoint will be prompted, which is OK and is the exact desired behavior.
• a new expname is specified and hence a new experiment directory would be used, since we do not want to concatenate and mix the new logging stats with the old ones.

Testing

Free view port rendering

• To render with camera2world matrices interpolated from the learned pose

python vis/free_viewport_rendering.py --load_dir /path/to/pretrained/exp_dir --render_type interpolate

• To render with spiral camera paths as in the original NeRF repo

python vis/free_viewport_rendering.py --load_dir /path/to/pretrained/exp_dir --render_type spiral

Visualize learned camera pose

python vis/plot_camera_pose.py --load_dir /path/to/pretrained/exp_dir

Notice that the learned camera phi & t is actually for camera2world matrices, the inverse of camera extrinsics

You will get a matplotlib window like this:

Road-map & updates

Basic NeRF model

• 2021-04-17 Basic implementation of the original paper, including training & logging
  • Add quaternion, axis-angle, rotation6D as the rotation representation
  • Add exp, square, ratio for different parameterizations of camera focal_x and focal_y
  • Siren-ized NeRF
  • refinement process as in the original NeRF-- paper
• Change DataParallel to DistributedDataParallel

Efficiency & training

• 2021-04-19 Add pre-train for 1000 epochs then refine for 10000 epochs, similar with the official code base.

• tolist - recent works in speeding up NeRF

More experiments

• vSLAM tasks and datasets
• traditional SfM datasets
• 2021-04-15 raw videos handling

Better SfM strategy

• tolist

More applicable for more scenes

• NeRF++ & NeRF-- for handling unconstrained scenes

• NeRF-W

• some dynamic NeRF framework for dynamic scenes

• Finish perceptual loss, for fewer shots

Related/used code bases

• GRAF ⬅️ The configuration, logging and checkpoints scripts is borrowed and modified from here.
• lucidrains' implementation of pi-GAN
• SIREN

Citations

• NeRF--

@article{wang2021nerf,
  title={Ne{RF}$--$: Neural Radiance Fields Without Known Camera Parameters},
  author={Wang, Zirui and Wu, Shangzhe and Xie, Weidi and Chen, Min and Prisacariu, Victor Adrian},
  journal={arXiv preprint arXiv:2102.07064},
  year={2021}
}

• SIREN

@inproceedings{sitzmann2020siren,
  author={Sitzmann, Vincent and Martel, Julien NP and Bergman, Alexander W and Lindell, David B and Wetzstein, Gordon},
  title={Implicit neural representations with periodic activation functions},
  booktitle={Proc. NeurIPS},
  year={2020}
}

• Perceptual model / semantic consistency from DietNeRF

@article{jain2021dietnerf,
  title={Putting NeRF on a Diet: Semantically Consistent Few-Shot View Synthesis},
  author={Ajay Jain and Matthew Tancik and Pieter Abbeel},
  journal={arXiv},
  year={2021}
}