New AI Identifies Burned Objects to Speed Up Fire Investigations and Insurance Claims

When a fire tears through a home or warehouse, what's left behind often barely resembles what was there before. A laptop becomes a melted slab. A couch becomes charred wire and foam. Researchers at the University of Illinois and collaborators point out that today's best AI vision systems handle blurry or dark photos fine, but they fall apart when the object itself has physically changed. Identifying those remnants matters for finding hazards, figuring out what was in a building, and settling insurance claims.

To study the problem, the team built TRACE, a test set of about 21,400 synthetic scenes built on real photographs, covering 499 objects across 189 categories — each shown both in perfect condition and at five increasing levels of fire damage. They then measured how well existing AI models performed. The results were stark: a popular detection model lost 71% of its accuracy from the mildest to the worst damage, and one large vision-language model's ability to match a burned object to its original form crashed from 93.85% down to 28.11%.

Their fix is called the Feature Recovery Module (FRM). In plain terms, it's a small add-on that sits inside an existing AI system and translates 'this is what a burned object looks like' back into 'this is what it looked like before.' The host AI stays frozen and untouched, so plugging it in is cheap and doesn't require retraining anything. Better yet, the worse the damage, the bigger the improvement it delivers.

Across different AI models and damage levels, the module boosted retrieval of the original object by about 12.5%, material identification by 20.1%, written descriptions by 13.2%, and reasoning about what the object was used for by 12.4%. The takeaway for non-engineers: AI is getting better at seeing through destruction, which could mean faster fire investigations, more accurate loss inventories, and fewer disputes over what was actually lost.