Feature Recovery for Object Understanding After Irreversible Fire Damage
TRACE tests whether vision models can recognize what burned objects used to be.
The paper introduces a benchmark with 21.4K synthetic post-fire scenes grounded in real images, covering 499 object identities across 189 categories. It measures detection, pristine-state retrieval, material recovery, description generation, and functional reasoning after damage changes an object’s actual structure. Existing models fall off sharply as damage worsens, including a 71% relative mAP drop for RF-DETR and InternVL3.5 retrieval R@1 falling from 93.85 to 28.11. The authors propose a frozen-host Feature Recovery Module that aligns degraded features with pristine ones and reports larger gains under more severe degradation. HF Daily Papers' note
The paper introduces a benchmark with 21.4K synthetic post-fire scenes grounded in real images, covering 499 object identities across 189 categories. It measures detection, pristine-state retrieval, material recovery, description generation, and functional reasoning after damage changes an object’s actual structure. Existing models fall off sharply as damage worsens, including a 71% relative mAP drop for RF-DETR and InternVL3.5 retrieval R@1 falling from 93.85 to 28.11. The authors propose a frozen-host Feature Recovery Module that aligns degraded features with pristine ones and reports larger gains under more severe degradation. HF Daily Papers' note
score 4