Feature Recovery for Object Understanding After Irreversible Fire Damage
arXiv:2609.12078 [cs.CV] ยท Submitted 10 September 2026
Authors: Aditi Tiwari, Sofia Stoica, Savya Khosla, David Forsyth, Heng Ji
Subjects: Computer Vision and Pattern Recognition (cs.CV) ยท ACM Classes: I.2.10; I.4.8
Overview
Objects in post-fire environments often undergo irreversible physical transformations that alter 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.
The TRACE Benchmark
To study this setting, the authors introduce TRACE, a transformation-aware benchmark for post-fire object understanding. TRACE comprises:
- 21.4K real-image-grounded synthetic scenes
- Paired object-level pristine-to-degraded progressions
- 499 object identities across 189 categories
The benchmark defines five tasks targeting localization and pre-degradation understanding:
- Degraded-object detection
- Pristine-state recovery and retrieval
- Original material recovery
- Pristine description generation
- Functional reasoning
- RF-DETR mAP decreases by 71% (relative)
- InternVL3.5 retrieval R@1 falls from 93.85 to 28.11
- Scene-level detection
- CLIP/SigLIP2 feature recovery
- All four object-level VLM tasks
- Comments: 28 pages, 11 figures, 9 tables
- Cite as: arXiv:2609.12078 [cs.CV]
- Version: arXiv:2609.12078v1
The Degradation Challenge
Existing models degrade sharply as fire severity increases. From the least to the most severe level:
Proposed Method: Feature Recovery Module (FRM)
To address this, the authors propose the Feature Recovery Module (FRM), a plug-and-play module that maps degraded encoder features to pristine-aligned representations while keeping the host model frozen. Trained only with paired feature supervision, FRM improves:
Gains are larger under more severe degradation. Across VLM hosts and severity levels, relative gains average:
| Task | Average Relative Gain |
|---|---|
| Retrieval | 12.5% |
| Material recovery | 20.1% |
| Description generation | 13.2% |
| Functional reasoning | 12.4% |
Paper Details
via ArXiv CV