This is the official repository for the paper "EAP-GS: Efficient Augmentation of Pointcloud for 3D Gaussian Splatting in Few-shot Scene Reconstruction" 🚀CVPR 2025 Nashville🚀
Before getting started, we outline the requirements for successfully running this code. The optimizer uses PyTorch and CUDA extensions in a Python environment.
- All of the requirements specified by 3DGS.
- COLMAP, which is a classical Structure-from-Motion (SfM) method.
- (optional) DetectorfreeSfM, which is another SfM method used in our work, leverages a detector-free matcher to enhances feature extraction in the texture-poor scenarios.
The repository contains submodules, clone this repository by:
git clone https://github.com/Osierddr/EAP-GS.git --recursiveOur default installation method is based on Conda package and environment management, allowing you to install the dependencies by:
conda env create --file environment.yml
conda activate EAP-GSIn this work, we select the train/test set in LLFF and Mip-NeRF360 datasets. Of course, you can also use your own captured datasets. Following the conventions of sparse-view settings, Camera poses are assumed to be known based on full-view estimation or other methods, located in the directory corresponding to the dataset. Our experiments use community standards train-test split, i.e., select every eighth image as the testing view, and evenly sample sparse views from the remaining images for training.
Download and place the dataset in <datadir>, copy the corresponding scenario's .json file from the /split directory into <datadir>, and prepare as follows:
<datadir>
|---<scene1>
| |---images
| | |---...
| |---sparse
| | |---0
| | |---...
| |---poses_bounds.npy
| |---split_index.json
|---<scene2>
|...
Run run_colmap.py, here we use the parameters set by FSGS to execute feature extraction, feature matching, and triangulation under known camera poses:
python run_colmap.py --base_path <datadir>/ --views 3 # 3 for LLFF, 12 for MipNeRF360The corresponding few-shot initialization results are in 3_views, and <points3D_3views.bin> is also the few-shot 3D pointcloud, your file path in <datadir> should be:
<datadir>
|---<scene1>
| |---3_views
| | |---created
| | |---images
| | |---sparse
| | |---database.db
| |---images
| | |---...
| |---sparse
| | |---0
| | |---points3D_3views.bin
| | |---...
| |---poses_bounds.npy
| |---split_index.json
|---<scene2>
|...
Run augmentation.py to get attention regions, which are placed under <datadir>/images with the same naming convention as images:
python augmentation.py --source_path <datadir>/<scene> --pc_name points3D_3viewsCommand Line Arguments for augmentation.py
Path to the source directory containing images.
Name of the pointcloud generated by SfM.. (points3D by default).
The maximum distance between two samples for one to be considered as in the neighborhood of the other. (15 by default).
The number of samples in a neighborhood for a point to be considered as a core point. (10 by default).
The radius of region around attention point. (10 by default).
Combine images and attention regions with same poses, running run_colmap.py again to generate a fine pointcloud, which is alse located in <datadir>/<scene>/sparse/0:
python run_colmap.py --base_path <datadir>/ --augment --views 3 # 3 for LLFF, 12 for MipNeRF360(optional) using DetectorfreeSfM can yields a more numerous and balanced pointcloud, which is adopted in our work.
Replace the previously generated coarse pointcloud with a fine pointcloud for subsequent reconstruction.
For training, we optimize for 5000 iterations, simply use:
python train.py -s <datadir>/<scene> --model_path <datadir>/<scene>/output -r 4 --seed 3 --pc_name points3D_3views_aug # Train without test set
python train.py -s <datadir>/<scene> --model_path <datadir>/<scene>/output -r 4 --seed 3 --pc_name points3D_3views_aug --eval # Train with train/test setAdditionally, we could introduce depth regularization by adding --depth and --usedepthReg, which are the same as DRGS:
python train.py -s <datadir>/<scene> --model_path <datadir>/<scene>/output -r 4 --seed 3 --pc_name points3D_3views_aug --eval --depth --usedepthRegOur code is built upon the foundation of following works, and we would like to thank the team for their valuable contributions!
Consider citing as below if you find this repository helpful to your project:
coming soon!