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, targeting surgical robotic perception. Developed by Falong Fan, the work was posted to arXiv (2603.24577) in March 2026. Soft-tissue reconstruction is challenging because low-texture surfaces, specular highlights, and instrument occlusions frequently fragment geometric continuity. To address this, EndoVGGT introduces a Deformation-aware Graph Attention (DeGAT) module that explicitly models tissue deformation, aiming to maintain consistent geometry across frames. The approach builds on vision-geometry-transformer-style architectures with a graph attention mechanism tailored to endoscopic surgical scenes, where tissues move and deform during procedures. This makes it relevant for downstream surgical robotics tasks such as scene understanding, tracking, and intraoperative guidance. The forum post provides a brief bilingual summary of the paper's abstract; full experimental details and benchmarks are available in the original arXiv paper.

Paper Overview

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

Abstract (translated from the original)

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. The authors propose EndoVGGT, a geometry-centric framework equipped with a Deformation-aware Graph Attention (DeGAT) module, designed to preserve consistent geometry in endoscopic surgical scenes.

Key Points

Problem: Reconstructing deformable soft tissue in 3D is difficult due to textureless surfaces, specular reflections, and occlusions caused by surgical instruments.
Approach: EndoVGGT augments a vision-geometry-transformer-style pipeline with a DeGAT module that explicitly accounts for tissue deformation, improving geometric consistency.
Application: Surgical robotic perception, including intraoperative scene understanding and guidance.

For full methodology, experiments, and benchmarks, see the original paper on arXiv.