火灾后物体理解的特征恢复基准与方法
原标题：Feature Recovery for Object Understanding After Irreversible Fire Damage
AI 摘要
研究者提出 TRACE，一个面向火灾后物体理解的变换感知基准，包含 21.4K 真实图像衍生的合成火灾场景和 499 个物体从完好到逐步退化的配对轨迹，覆盖 189 个类别。实验显示现有检测器、编码器和 VLM 在物理退化下性能大幅下降，如 RF-DETR mAP 相对下降 71%，InternVL3.5 检索 R@1 从 93.85 降至 28.11。为此作者提出轻量即插即用的 Feature Recovery Module（FRM），在冻结宿主模型的情况下将退化特征映射到完好对齐表示，在检测、特征恢复和四项 VLM 任务上均取得提升。
正文节选
Feature Recovery for Object Understanding After Irreversible Fire Damage Abstract Objects in post-fire environments often undergo irreversible physical transformations that change their geometry, material state, and visual appearance. Detecting and identifying these remnants is critical for locating hazards, reconstructing pre-incident contents, and inventorying losses. Unlike standard image corruptions, these degradations affect the physical structure of the object itself. To study this setting