Gaussian Splatting on the Move:
Blur and Rolling Shutter Compensation for Natural Camera Motion

Installation

Prerequisites: run on a Linux system with a recent NVidia RTX GPU with at least 8 GB of VRAM. Git must be installed.

1. Activate a Conda environment with PyTorch that supports Nerfstudio
2. Possibly required, depending on your environment: conda install -c conda-forge gcc=12.1.0
3. Run ./scripts/install.sh (see steps within if something goes wrong)

Training with custom data

Custom video data (new in version 2): The method can now be used for motion blur compensation with plain video data as follows

./scripts/process_and_train_video.sh /path/to/video.mp4

or for rolling shutter compensation as

ROLLING_SHUTTER=ON ./scripts/process_and_train_video.sh /path/to/video.mp4

Currently simultaneous motion blur and rolling-shutter compensation is only possible with known readout and exposure times. The easiest way to achieve this is using the Spectacular Rec application to record the data (see below).

Spectacular Rec app (v1.0.0+ for Android, v1.2.0+ for iOS) is needed for simultaneous rolling shutter and motion blur compensation. This approach is also expected to give the best results if the data collection app can be chosen, since it also allows automatic blurry frame filtering and VIO-based velocity initialization, both of which improve the final reconstruction quality. Instructions below.

First, download and extract a recording created using the app, e.g., /PATH/TO/spectacular-rec-MY_RECORDING.

Then process as

./scrpts/process_and_train_sai_custom.sh /PATH/TO/spectacular-rec-MY_RECORDING

or, for a faster version:

SKIP_COLMAP=ON ./scrpts/process_and_train_sai_custom.sh /PATH/TO/spectacular-rec-MY_RECORDING

See the contents of the script for more details.

Comparison videos To train a custom recording with and without motion blur compensation and render a video comparing the two, use this script:

• motion blur OR rolling shutter, COLMAP-based, from video:

   ./scripts/render_and_compile_comparison_video.sh /path/to/video.mp4
   ROLLING_SHUTTER=ON ./scripts/render_and_compile_comparison_video.sh /path/to/video.mp4
  

• motion blur AND rolling shutter compensations (needs Spectacular Rec data)

   ./scripts/render_and_train_comparison_sai_custom.sh /PATH/TO/spectacular-rec-MY_RECORDING
  

Benchmark data

The inputs directly trainable with our fork of Nerfstudio are stored in data/inputs-processed folder. Its subfolders are called "datasets" in these scripts.

The data can be automatically downloaded by first installing: pip install unzip and then running

python download_data.py --dataset synthetic
# or 'sai' for processed real world smartphone data

The data folder structure is as follows:

<3dgs-deblur>
|---data
    |---inputs-processed
        |---colmap-sai-cli-vels-blur-scored/
            |---iphone-lego1
                |---images
                    |---image 0
                    |---image 1
                    |---...
                |---sparse_pc.ply
                |---transforms.json
            |---...
        |---synthetic-mb
            |---cozyroom
                |---images
                    |---image 0
                    |---image 1
                    |---...
                |---sparse_pc.ply
                |---transforms.json
            |---...
        |---...
|---...

Training

Example: List trainable variants for the synthetic-mb dataset:

python train.py --dataset=synthetic-mb

Train a single variant

python train.py --dataset=synthetic-mb --case=2

Common useful options:

• --dry_run
• --preview (show Viser during training)

Additionally, any folder of the form data/inputs-processed/CASE can be trained directly with Nerfstudio using the ns-train splatfacto --data data/inputs-processed/CASE .... Use --help and see train.py for the recommended parameters.

Viewing the results

Results are written to data/outputs/ by dataset. You can also run these on another machine and download these results on your machine. All of the below commands should then work for locally examining the results.

Numeric

List all numeric results

python parse_outputs.py

... or export to CSV

python parse_outputs.py -f csv > data/results.csv

Visualizations

Off-the-shelf:

• Viser: ns-viewer --load-config outputs/DATASET/VARIANT/splatfacto/TIMESTAMP/config.yml (show actual results)
• Tensorboard: tensorboard --logdir outputs/DATASET/VARIANT/splatfacto/TIMESTAMP (prerequisite pip install tensorboard)

Custom:

• Created by train.py --render_images ...: Renders of evaluation images and predictions are available in outputs/DATASET/VARIANT/splatfacto/TIMESTAMP (/renders, or /demo_video*.mp4 if render_video.py has been run, see below)
• Demo videos: see render_video.py and scripts/render_and_combine_comparison_video.sh

Processing the raw benchmark input data

This method also creates the extra variants discussed in the appendix/supplementary material of the paper, as well as all the relevant synthetic data variants.

Synthetic data

For synthetic data, we use different re-rendered versions of the Deblur-NeRF synthetic dataset. Note that there exists several, slightly different variation, which need to be trained with correct parameters for optimal results.

Our Deblur-NeRF re-render (uses $\gamma = 2.2$): Download and process as:

python download_data.py --dataset synthetic-raw
python process_synthetic_inputs.py

Other variants

1. Download the data and extract as inputs-raw/FOLDER_NAME (see options below)

2. Run

   python process_deblur_nerf_inputs.py --dataset=FOLDER_NAME --manual_point_cloud all
   

This creates a dataset called colmap-DATASET-synthetic-novel-view-manual-pc Note that it may be necessary to run the last command multiple times until COLMAP succeeds in all cases (see also the --case=N argument in the script).

Supported datasets (TODO: a bit messy):

• Original Deblur-NeRF: FOLDER_NAME = synthetic_camera_motion_blur. Uses $\gamma = 2.2$.
• BAD-NeRF re-render: FOLDER_NAME = nerf_llff_data. Uses $\gamma = 1$.
• BAD-Gaussians re-render: FOLDER_NAME = bad-nerf-gtK-colmap-nvs

The last two are very similar except for the "Tanabata" scene, which is broken in the BAD-NeRF version: the underlying 3D model is slightly different in the (sharp) and training (blurry) images (objects moved around).

Smartphone data

Download as:

python download_data.py --dataset sai-raw

and then process and convert using the following script:

./scripts/process_smartphone_dataset.sh
# or 
# EXTRA_VARIANTS=ON ./scripts/process_smartphone_dataset.sh

Note: all the components in this pipeline are not guaranteed to be deterministic, especially when executed on different machines. Especially the COLMAP has a high level of randomness.

Changelog

Version 2 (2024-05)

• Angular and linear velocities added as optimizable variables, which can be initialized to zero if VIO-estimated velocity data is not available (i.e., no IMU data available)
• Added --optimize-eval-cameras mode, which allows optimizing evaluation camera poses and velocities (if --optimize-eval-velocities=True) without back-propagating information to the 3DSG reconstruction. This replaces the previous two-phase optimization mode (called "rolling shutter pose optimization" in the first paper revision)
• Method can be run in motion blur OR rolling-shutter mode form plain video without a known exposure or readout times. Added a helper script process_and_train_video.sh for this.
• Rebased on Nerfstudio version 1.1.0 and gsplat 409bcd3c (based on 0.1.11)
• Fixed a bug in pixel velocity formulas
• Tuned hyper-parameters (separate parameters for synthetic and real data)
• Using optimizable background color by KevinXu02 for synthetic data
• Using $\gamma \neq 1$ and --min-rgb-level only when motion blur compensation is enabled (for a more fair comparison to Splatfacto)
• Added conversion scripts for other common Deblur-NeRF dataset variants

Version 1 (2024-03)

Initial release where IMU data was mandatory to run the method, and the uncertainties in VIO-estimated velocities were addressed with a custom regularization scheme (see §3.6 in the first revision of the paper). Based on Nerfstudio version 1.0.2 and gsplat 0.1.8.

License

The code in this repository (except the gh-pages website branch) is licensed under Apache 2.0. See LICENSE and NOTICE files for more information.

For the source code of the website and its license, see the gh-pages branch.

The licenses of the datasets (CC BY-SA 4.0 & CC BY 4.0) are detailed on the Zenodo pages.