GeBDA: Building Damage Assessment as Text-Based Sequence Prediction with Vision-Language Models

Forum topic · 小凯 · 2026-09-01

Summary

This paper (arXiv:2608.28567) by Olivier Dietrich, Krishna Sapkota, Konrad Schindler, and Genady Beryozkin explores whether general-purpose Vision-Language Models (VLMs) can perform Building Damage Assessment (BDA) through autoregressive sequence generation alone, without dedicated network architectures or fine-tuned geospatial foundation models. The authors reformulate BDA as the task of predicting a variable-length set of bounding boxes, where each box is specified by its coordinates and a damage label, generated as text. Their preliminary implementation builds on the open Gemma model and achieves promising damage mapping results using only bi-temporal satellite images and a suitable text prompt. The work suggests that general VLMs can serve as flexible tools for geospatial disaster-response analysis via pure text-based sequence prediction.

Paper Overview

Field: Computer Vision (CV)
Authors: Olivier Dietrich, Krishna Sapkota, Konrad Schindler, Genady Beryozkin
Published: 2026-08-28
arXiv: 2608.28567

Original Abstract

Conventionally, Building Damage Assessment (BDA) is tackled either with dedicated network architectures or by fine-tuning geospatial image foundation models. In this work, we ask whether a general-purpose Vision-Language Model (VLM) can localize buildings and grade their damage through autoregressive sequence generation alone. We cast BDA as predicting a variable-length set of bounding boxes, each specified by its coordinates and a damage label. Our preliminary implementation, based on the open Gemma model, achieves promising damage mapping results from only bi-temporal satellite images and a suitable text prompt.

Key Points

New formulation: BDA is recast as a text-based sequence prediction task — the model outputs a variable-length set of bounding boxes, each defined by coordinates plus a damage label.
No specialized architecture: Instead of dedicated damage-assessment networks or fine-tuned geospatial foundation models, a general-purpose VLM handles the task end-to-end via autoregressive generation.
Implementation: The preliminary version is built on the open Gemma model.
Inputs: Only bi-temporal (pre- and post-disaster) satellite images together with a suitable text prompt.
Results: Promising damage mapping results, indicating VLMs can localize buildings and grade damage without task-specific training.

Why It Matters

Treating geospatial damage mapping as plain text generation simplifies the pipeline and suggests that general vision-language models could become flexible, prompt-driven tools for disaster response and remote sensing analysis.

--- *Auto-collected on 2026-09-01*