Chong Cheng^1* · Yu Hu^1* · Sicheng Yu¹ · Beizhen Zhao¹ · Zijian Wang¹ · Hao Wang^1†

International Conference on Computer Vision, ICCV 2025

¹The Hong Kong University of Science and Technology (Guangzhou)

The code has been tested on systems with:

• Ubuntu 22.04 LTS
• Python 3.10.18
• CUDA 11.8
• NVIDIA GeForce RTX 3090 or A6000

Clone the repo with --recursive because we have submodules:

git clone https://github.com/3DAgentWorld/RegGS.git --recursive
cd RegGS

This codebase has been successfully tested with Python 3.10, CUDA 11.8, and PyTorch 2.5.1. We recommend installing the dependencies in a virtual environment such as Anaconda.

1. Install main libraries:

   conda env create -f environment.yaml
   
   conda activate reggs
   
   pip install -r requirements.txt

2. Install thirdparty submodules:

   pip install thirdparty/diff-gaussian-rasterization-w-pose
   
   pip install thirdparty/gaussian_rasterizer`
   
   pip install thirdparty/simple-knn

3. Compile the cuda kernels for RoPE (as in CroCo v2):

   cd src/noposplat/model/encoder/backbone/croco/curope
   python setup.py build_ext --inplace

4. If you encounter cannot import torch. add option --no-build-isolation to pip install

Download NoPoSplat re10k checkpoints and acid checkpoints to ./pretrained_weights directory, run:

wget -c https://huggingface.co/botaoye/NoPoSplat/resolve/main/re10k.ckpt -P ./pretrained_weights

wget -c https://huggingface.co/botaoye/NoPoSplat/resolve/main/acid.ckpt -P ./pretrained_weights

The official RE10K dataset can be downloaded from: https://google.github.io/realestate10k/download.html

The preprocessed re10k data is placed in the directory ./sample_data. To run RegGS on sample data, run:

1. The inference stage:
   CUDA_VISIBLE_DEVICES=0 python3 run_infer.py config/re10k.yaml
2. The refinement stage:
   CUDA_VISIBLE_DEVICES=0 python3 run_refine.py --checkpoint_path output/re10k/000c3ab189999a83
3. The evaluation stage:
   CUDA_VISIBLE_DEVICES=0 python3 run_metric.py --checkpoint_path output/re10k/000c3ab189999a83

To convert raw RE10K trajectory files to RegGS sample format, use the preprocessing script:

python scripts/preprocess_re10k.py \
  --scene-txt datasets/RealEstate10K/test/SCENE_ID.txt \
  --original-width 455 \
  --target-width 256 \
  --download-video

Parameters:

• --scene-txt: Path to RE10K scene txt file (required)
• --out-root: Output root directory (default: sample_data)
• --original-width: Original frame width from video (e.g., 455) for intrinsics adjustment
• --target-width: Target frame width after cropping (default: 256)
• --video-path: Path to local video file (mp4)
• --download-video: Download video using yt-dlp if --video-path not provided
• --max-frames: Maximum number of frames to process (default: -1, meaning all frames)
• --overwrite: Overwrite existing output directory

The script outputs:

• images/: PNG frames extracted from video
• cameras.json: Per-frame camera poses and intrinsics in RegGS format
• intrinsics.json: Normalized camera intrinsics parameters

The ACID (Airborne Coastal Imagery Dataset) can be downloaded from: https://infinite-nature.github.io/

To convert raw ACID trajectory files to RegGS sample format, use:

python scripts/preprocess_acid.py \
  --scene-txt datasets/acid/test/SCENE_ID.txt \
  --original-width 455 \
  --target-width 256 \
  --download-video

Parameters:

• --scene-txt: Path to ACID scene txt file (required)
• --out-root: Output root directory (default: sample_data)
• --original-width: Original frame width before 256 preprocessing (default: 455)
• --target-width: Target frame width after preprocessing (default: 256)
• --video-path: Path to local video file (mp4)
• --download-video: Download video using yt-dlp if --video-path is not provided
• --max-frames: Maximum number of frames to process (default: -1, meaning all frames)
• --overwrite: Overwrite existing output directory

The script outputs:

• images/: PNG frames extracted from video
• cameras.json: Per-frame camera poses and intrinsics in RegGS format
• intrinsics.json: Normalized camera intrinsics parameters

• create codebase
• add evaluation script
• prepare sample data
• write installation guide
• add data preprocessing script
• implement GPU-optimized k-means
• add Gradio visualization

@inproceedings{cc2025_reggs,
  title     = {{RegGS}: Unposed Sparse Views Gaussian Splatting with {3DGS} Registration},
  author    = {Cheng, Chong and Hu, Yu and Yu, Sicheng and Zhao, Beizhen and Wang, Zijian and Wang, Hao},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year      = {2025}
}