uva-cv-lab/CLICK3R

CLICK3R: Interactive Stereo 3D Reconstruction with Sparse Correspondence Clicks

CLICK3R is a framework for interactive geometric correction of feed-forward models using sparse human correspondence clicks. Given an image pair and a small set of cross-view point correspondences provided by a user, CLICK3R injects geometric constraints into the 3D reconstruction networks through a lightweight point correspondence adapter, enabling the model to resolve ambiguities in challenging scenarios such as repetitive textures, symmetric structures, large viewpoint changes, or textureless regions.

Model Zoo

Model	Description	Link
CLICK3R	ViT-L encoder, ViT-B decoder, 566.6M params, 224 × 224. The model of the paper.	Download
DUSt3R (retrained)	Same encoder and decoder, 532.3M params, 224 × 224. DUSt3R trained on the same data as CLICK3R, the baseline of the paper.	Download

Usage

Run the interactive demo. The launcher sets up the environment and downloads the checkpoint:

git clone https://github.com/UVA-Computer-Vision-Lab/click3r
cd click3r
python demo.py

Download a checkpoint directly:

hf download uva-cv-lab/CLICK3R click3r.pth --local-dir checkpoints
hf download uva-cv-lab/CLICK3R-DUSt3R-retrained retrained.pth --local-dir checkpoints/dust3r

Evaluation and training commands are in the code repository.

Citation

@inproceedings{gu2026click3r,
  title={CLICK3R: Interactive Stereo 3D Reconstruction with Sparse Correspondence Clicks},
  author={Gu, Hao and Yao, Jin and Cheng, Zezhou},
  booktitle={Neural Information Processing Systems (NeurIPS)},
  year={2026}
}

License

The model checkpoints are released under CC BY-NC-SA 4.0, as they are initialized from CroCo v2 and trained on non-commercial datasets.