Paper Overview
Field: Computer Vision Authors: Hao Gu, Jin Yao, Zezhou Cheng Published: 2026-11-09 arXiv: 2611.12875
Abstract (Translated)
Despite significant progress in feed-forward large 3D reconstruction models, they produce a single, irreversible prediction — once the model fails, there is no mechanism for correction. To address this, the authors propose Click3R, a framework that performs interactive geometric correction of feed-forward models using sparse human correspondence clicks. Given a pair of images and a small number of user-provided cross-view correspondences, Click3R injects geometric constraints into the 3D reconstruction network through a lightweight point correspondence adapter. This enables the model to resolve ambiguities in challenging scenarios such as repetitive textures, symmetric structures, large viewpoint changes, or textureless regions.
To evaluate interactive reconstruction, the authors construct a challenging benchmark specifically designed to expose the failure modes of existing feed-forward methods. Experiments show that Click3R reduces reconstruction error with just a single correspondence click, achieving improvements on both standard benchmarks and the newly proposed benchmark. The results highlight that sparse human interaction provides an effective and practical route for correcting geometric reconstruction errors.
Original Abstract (Excerpt)
> We introduce Click3R, a framework for interactive geometric correction of feed-forward 3D reconstruction models using sparse human correspondence clicks. Click3R injects geometric constraints into 3D reconstruction networks through a lightweight point correspondence adapter, enabling correction of errors in challenging scenarios.
---
*Auto-collected on 2026-11-11.*