Click3R / README.md
yaojin17's picture
Link the arXiv paper
1a236e0 verified
|
Raw History Blame Contribute Delete
4.69 kB

A newer version of the Gradio SDK is available: 6.32.0

Upgrade
metadata
title: Click3R
emoji: 📐
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 6.31.0
python_version: '3.12'
app_file: app.py
pinned: false
license: cc-by-nc-sa-4.0
short_description: Interactive stereo 3D reconstruction with clicks
models:
  - uva-cv-lab/Click3R
  - naver/DUSt3R_ViTLarge_BaseDecoder_224_linear

Click3R — Gradio Space

Demo for Click3R: Interactive Stereo 3D Reconstruction with Sparse Correspondence Clicks (NeurIPS 2026).

On first launch the Space downloads the checkpoint from uva-cv-lab/Click3R via huggingface_hub.hf_hub_download, and the DUSt3R comparison weights from naver/DUSt3R_ViTLarge_BaseDecoder_224_linear. Paper: arXiv:2611.12875. Code: UVA-Computer-Vision-Lab/click3r.

Examples

The demo has 23 example pairs. Errors are relative pose errors in degrees.

From the paper. Bicycle (Figure 1, CO3D) and Railing (Figure 4, BlendedMVS). Their clicks are transcribed from the figures.

Human clicks. Pairs from Hard3DBench with clicks from a human annotator. Errors are rotation errors.

Example Click pairs DUSt3R Click3R with clicks
Breakfast room 2 175.7° 9.2°
Relief 4 119.6° 9.3°
Scribbled box 3 87.4° 1.9°
Brick building model 3 84.9° 5.8°

Automatically selected. We searched about 15,800 test pairs for ones where pretrained DUSt3R errs by at least 15° and Click3R without clicks by at least 12°, and where Click3R with at most four clicks stays below 12°, gains at least 8° and stays below 20° when the clicks move by up to 2 pixels. We then picked these by hand from the 372 that qualified, for a visible change in the 3D result. Their clicks are ground-truth correspondences at corners visible in both views, not human clicks. Errors are the larger of the rotation and translation errors.

Example Click pairs DUSt3R Click3R without clicks Click3R with clicks
Cellphone 1 165.0° 133.5° 3.1°
Suitcase 1 175.6° 111.7° 2.7°
Toaster 1 176.1° 163.9° 4.0°
Toy truck 1 170.4° 161.7° 3.4°
Umbrella 1 74.1° 161.7° 3.1°
Cake 1 40.7° 38.9° 3.6°
Teddy bear 2 124.1° 127.3° 2.6°
Handbag 3 73.1° 172.5° 3.9°
Building model, top view 1 135.5° 78.4° 8.8°
Poster box 4 67.0° 98.9° 4.4°
Rabbit figurine with ear muffs 4 163.1° 172.5° 3.3°
Aerial houses 1 75.1° 178.8° 6.1°
Town rooftops 2 93.1° 133.0° 2.8°
Village and pond 1 65.3° 177.6° 0.6°
Smartphone 2 180.0° 179.1° 3.7°
Flower pot 1 67.7° 72.4° 8.0°
Courtyard 1 88.1° 89.4° 1.3°

Example images

The images have been converted to PNG and/or center-cropped and resized for the demo. Dataset licenses apply to these derived images. No endorsement by the dataset creators is implied.

Example Attribution Dataset/asset terms and source
Bicycle, Cellphone, Suitcase, Toaster, Toy truck, Umbrella, Cake, Teddy bear, Handbag, Smartphone Meta and CO3D contributors; Reizenstein et al., Common Objects in 3D, ICCV 2021 CO3D, CC BY-NC 4.0
Railing, Aerial houses, Town rooftops, Village and pond Yao et al., BlendedMVS: A Large-scale Dataset for Generalized Multi-view Stereo Networks, CVPR 2020 BlendedMVS, CC BY 4.0
Breakfast room Wig42, scene creator; Azinović et al., Neural RGB-D Surface Reconstruction, CVPR 2022, rendered dataset Original scene, CC BY; dataset attribution
Relief, Courtyard Schöps et al., A Multi-View Stereo Benchmark with High-Resolution Images and Multi-Camera Videos, CVPR 2017 ETH3D, CC BY-NC-SA 4.0
Scribbled box; Brick building model; Building model, top view; Poster box; Rabbit figurine with ear muffs Jensen et al., Large Scale Multi-view Stereopsis Evaluation, CVPR 2014 DTU Robot Image Data, citeware
Flower pot Xia et al., RGBD Objects in the Wild: Scaling Real-World 3D Object Learning from RGB-D Videos, CVPR 2024 WildRGB-D, MIT