FLASH3R: On-Board Incremental Dense 3D Aerial Reconstruction with Progressive Geometry Guided Visual Geometry Grounded Transformer

Zhihao Jia¹, Han Hu*¹, Pengxin Gu¹, Junfan Wang¹, Yang Jia², Bo Xiang², Jiwei Deng³, Yulin Ding¹, Qing Zhu¹

¹ Faculty of Geosciences and Engineering, Southwest Jiaotong University · ² Sichuan Highway Planning, Survey, Design and Research Institute Ltd. · ³ China Railway Design Corporation

preprint · 2026

Paper Code

Background and Motivation

UAV photogrammetry is increasingly used for time-sensitive field tasks such as emergency response, disaster inspection, and infrastructure monitoring. In these scenarios, the basic requirement is not only to acquire 2D orthophotos, but to generate real-time, geo-referenced 3D information that can support deformation analysis, volume estimation, and rapid decision-making.

Conventional on-board pipelines are often limited to digital orthophoto maps (DOM), while full sparse-to-dense reconstruction remains computationally expensive for edge UAV platforms. Flash3R targets this gap by combining incremental SfM with feed-forward visual geometry inference, enabling dense 3D reconstruction directly on the UAV without depending on cloud computation.

Sparse-Dense Dual-Branch Parallel Reconstruction
Image-Level and Point-Level Dense-to-Sparse Alignment
Edge Device Deployment Validation

Pipeline

Flash3R processes successive UAV image batches through three compact stages:

Sparse Pipeline: geo-referenced SfM scaffold
Dense Pipeline: transformer-based dense geometry inference
Two-Level Alignment: image-level pose refinement and point-level MLS correction