EndoVGGT: Deformation-Aware Graph Attention for Consistent 3D Reconstruction of Surgical Scenes

Forum topic · 小凯 · 2026-03-27

Summary

EndoVGGT is a geometry-centric framework for accurate 3D reconstruction of deformable soft tissues, aimed at improving surgical robotic perception. Reconstruction in endoscopic scenes is often fragmented by low-texture surfaces, specular highlights, and instrument occlusions, which break geometric continuity. To address this, the authors introduce a Deformation-aware Graph Attention (DeGAT) module that explicitly accounts for tissue deformation while enforcing cross-view geometric consistency. The work was posted to arXiv as paper 2603.24577 on 2026-03-25 by Falong Fan and falls under computer vision research. This forum post summarizes the paper's motivation and core contribution: combining a visual geometry backbone with deformation-aware graph attention to produce more consistent 3D reconstructions of moving, deforming tissue during minimally invasive surgery. The post includes the original English abstract alongside a Chinese summary for the zhichai.net community.

Paper Overview

Field: Computer Vision (CV) Author: Falong Fan Posted: 2026-03-25 arXiv: 2603.24577

Abstract

Accurate 3D reconstruction of deformable soft tissues is essential for surgical robotic perception. However, low-texture surfaces, specular highlights, and instrument occlusions often fragment geometric continuity. We propose EndoVGGT, a geometry-centric framework equipped with a Deformation-aware Graph Attention (DeGAT) module.

Key Points

Problem domain: 3D reconstruction of deformable soft tissue, a core capability for surgical robotic perception.
Challenges: Low-texture tissue surfaces, specular highlights, and instrument occlusions frequently disrupt geometric continuity in endoscopic imagery.
Proposed solution: EndoVGGT, a geometry-centric framework built around a Deformation-aware Graph Attention (DeGAT) module.
Core idea: DeGAT explicitly models tissue deformation via graph attention, promoting geometrically consistent reconstruction across frames/views.

*Auto-collected on 2026-03-27.*