This is the coarse decomposition part of the method proposed in AutoRecon: Automated 3D Object Discovery and Reconstruction. It can be used to preprocess a casual capture (object-centric multi-view images or a video) which estimate the camera poses with SfM and localize the salient foreground object for further reconstruction.
Please install AutoDecomp following INSTALL.md.
Here we takes assets/custom_data_example/co3d_chair
as an example.
You can run automatic foreground scene decomposition with: scripts/run_pipeline_demo_low-res.sh
.
You should get a similar visualization as in assets/custom_data_example/co3d_chair/vis_decomposition.html
.
You can take the data structure and the script as a reference to run the pipeline on your own data.
- Download the demo data from Google Drive and put them under
data/
. - Run one of the script in
scripts/test_pipeline_co3d_manual-poses/cvpr
(use low-res images for feature matching and DINO features) orscripts/test_pipeline_co3d_manual-poses
(use high-res images for feature matching and DINO features) to run the inference pipeline. - We save camera poses, decomposition results and visualization to
path_to_the_instance/auto-deocomp_sfm-transformer
.
We also support import camera poses saved in the IDR format and localize the foreground object. You can run one of the script in scripts/test_pipeline_bmvs/cvpr
or scripts/test_pipeline_bmvs
for reference.
If you find this code useful for your research, please use the following BibTeX entry.
@inproceedings{wang2023autorecon,
title={AutoRecon: Automated 3D Object Discovery and Reconstruction},
author={Wang, Yuang and He, Xingyi and Peng, Sida and Lin, Haotong and Bao, Hujun and Zhou, Xiaowei},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={21382--21391},
year={2023}
}