# Self-Supervised Correspondence Estimation via Multiview Registration

**Authors:** Mohamed El Banani, Ignacio Rocco, David Novotny, Andrea Vedaldi, Natalia Neverova, Justin Johnson, Benjamin Graham  
**arXiv:** [2212.03236](https://arxiv.org/abs/2212.03236)  
**Date:** 2022-12-06  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2212.03236v1/

## Key points

- The paper addresses the issue of learning correspondence estimation from video sequences
- Previous approaches focus on spatio-temporal consistency from close-by frame pairs, but fail to capture long-range consistency between distant overlapping frames
- The authors propose a self-supervised approach that leverages multiview consistency in short RGB-D video sequences
- The method combines pairwise correspondence estimation and registration using a novel SE(3) transformation synchronization algorithm
- Self-supervised multiview registration allows for obtaining correspondences over longer time frames, increasing diversity and difficulty of sampled pairs
- Experiments were conducted on indoor scenes for correspondence estimation and RGB-D point cloud registration
- Results show that the proposed approach performs on par with supervised approaches, demonstrating its effectiveness in learning accurate correspondences.

## Summary

The paper titled "Self-Supervised Correspondence Estimation via Multiview Registration" addresses the issue of learning correspondence estimation from video sequences. Previous approaches have focused on utilizing spatio-temporal consistency from close-by frame pairs, but fail to capture the richer long-range consistency between distant overlapping frames. To overcome this limitation, the authors propose a self-supervised approach that leverages multiview consistency in short RGB-D video sequences. The proposed method combines pairwise correspondence estimation and registration using a novel SE(3) transformation synchronization algorithm. By incorporating self-supervised multiview registration, correspondences can be obtained over longer time frames, increasing both the diversity and difficulty of sampled pairs. This allows for more robust visual learning by capturing the full range of spatio-temporal relationships within a video sequence. To evaluate their method, experiments were conducted on indoor scenes for correspondence estimation and RGB-D point cloud registration. The results show that their approach performs on par with supervised approaches, demonstrating its effectiveness in learning accurate correspondences. In conclusion, this paper presents a novel self-supervised approach for correspondence estimation that takes advantage of multiview consistency in short RGB-D video sequences to improve upon existing approaches and achieve comparable performance to supervised methods in various tasks related to visual learning.