Flow Distillation Sampling: Regularizing 3D Gaussians with Pre-trained Matching Priors

Lin-Zhuo Chen^{1 *}  Kangjie Liu^{1 *}  Youtian Lin¹  Zhihao Li²  Siyu Zhu³  Xun Cao¹  Yao Yao^{1 †} 

This is official implement of our ICLR 2025 paper: Flow Distillation Sampling: Regularizing 3D Gaussians with Pre-trained Matching Priors.

📝 Abstract

3D Gaussian Splatting (3DGS) has achieved excellent rendering quality with fast training and rendering speed. However, its optimization process lacks explicit geometric constraints, leading to suboptimal geometric reconstruction in regions with sparse or no observational input views. In this work, we try to mitigate the issue by incorporating a pre-trained matching prior to the 3DGS optimization process. We introduce Flow Distillation Sampling (FDS), a technique that leverages pre-trained geometric knowledge to bolster the accuracy of the Gaussian radiance field. Our method employs a strategic sampling technique to target unobserved views adjacent to the input views, utilizing the optical flow calculated from the matching model (Prior Flow) to guide the flow analytically calculated from the 3DGS geometry (Radiance Flow). Comprehensive experiments in depth rendering, mesh reconstruction, and novel view synthesis showcase the significant advantages of FDS over state-of-the-art methods. Additionally, our interpretive experiments and analysis aim to shed light on the effects of FDS on geometric accuracy and rendering quality, potentially providing readers with insights into its performance.

🚀 Getting Started

Data preparation

1. Download our colmap points for 2DGS initilization: mushroom_colmap.
2. Download mushroom dataset: mushroom_website.
3. Put our colmap points into mushroom dataset:

FDS
├── Mushroom
    ├── activity
    |   ├── iphone
    |   ├── ├── long_capture
    |   ├── ├── ├── put colmap points here

    ├── classroom
    |   ├── iphone
    |   ├── ├── long_capture
    |   ├── ├── ├── put colmap points here

...

Installation

Clone FDS

git clone https://github.com/NJU-3DV/FDS.git --recursive

Install Pointrix

cd pointrix
git submodule update --init --recursive

Please refer to https://github.com/pointrix-project/Pointrix for the install instruction.

Running

Mushroom dataset

python launch.py --config configs/mushroom_config.yaml \
                  trainer.datapipeline.dataset.data_path=[your_data_path] \
                  trainer.output_path=[your_log_path] \
                  trainer.exporter.exporter_b.extra_cfg.gt_mesh_path=[your_mesh_path]  \
                  trainer.gui.viewer_port=8005

for example, to run vr room scene in mushroom dataset:

python launch.py --config configs/mushroom_config.yaml \
                        trainer.datapipeline.dataset.data_path=/NASdata/clz/data/mushroom/vr_room/iphone \
                        trainer.output_path=/NASdata/clz/log/fds_paper_final_v2/2dgs/fds_test/vr_room \
                        trainer.exporter.exporter_b.extra_cfg.gt_mesh_path=/NASdata/clz/data/mushroom/vr_room \
                        trainer.gui.viewer_port=8005

TODO

• More stable results
• DTU datasets.
• Supervised with more prior information.

Acknowledgements

Thanks to the following repos for their great work, which helps us a lot in the development of FDS:

• Pointrix: A light weight framework for gaussian points rendering.
• DN Splatter: The mesh exporter and mushroom dataloader.
• 2d gaussian splatting: Rendering Kernel.
• Sea Raft: Optical flow model of FDS.

Citation

If you find this work is useful for your research, please cite our paper:

@inproceedings{chen2024fds, 
        title={Flow Distillation Sampling: Regularizing 3D Gaussians with Pre-trained Matching Priors}, 
        author={Lin-Zhuo Chen and Kangjie Liu and Youtian Lin and Zhihao Li and Siyu Zhu and Xun Cao and Yao Yao}, 
        booktitle={ICLR}, 
        year={2025}
      }