TRACE Released: A New Benchmark for Post-Fire Object Understanding and Feature Recovery

Summary:The TRACE research team introduces TRACE, a novel benchmark for evaluating object understanding and feature recovery in post-fire environments. TRACE encompasses 21.4K real-image-grounded synthetic scenes and 499 object-level pristine-to-degraded progressions across 189 categories, targeting five key tasks: degraded-object detection, pristine-state recovery and retrieval, original material recovery, pristine description generation, and functional reasoning. To address the sharp performance degra

---

Overview

In post-fire environments, objects often undergo irreversible physical transformations that alter their geometry, material state, and visual appearance. Detecting and identifying these remnants is crucial for hazard localization, pre-incident content reconstruction, and loss inventory. The TRACE research team introduces TRACE, a novel benchmark designed to address the challenges of object understanding in such complex scenarios.

TRACE Benchmark

TRACE includes 21.4K real-image-grounded synthetic scenes and 499 object-level pristine-to-degraded progressions across 189 categories, targeting five core tasks:

• Degraded-Object Detection: Identifying the location and category of damaged objects.
• Pristine-State Recovery and Retrieval: Recovering the original state of objects and retrieving relevant information.
• Original Material Recovery: Identifying the material composition of objects before degradation.
• Pristine Description Generation: Generating detailed descriptions of objects in their original state.
• Functional Reasoning: Inferring the functional purpose of objects in their original state.

Limitations of Existing Models

Experiments show that existing models suffer from sharp performance degradation under severe degradation conditions. For instance, the mAP of RF-DETR drops by 71% from minor to severe degradation, while the retrieval R@1 of InternVL3.5 falls from 93.85 to 28.11.

Feature Recovery Module (FRM)

To address this, the team proposes 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 significantly enhances scene-level detection, CLIP/SigLIP2 feature recovery, and all four object-level VLM tasks. The advantages of FRM are more pronounced under more severe degradation conditions.

Performance Improvements

Across VLM hosts and degradation severity levels, FRM achieves average relative gains of 12.5% for retrieval, 20.1% for material recovery, 13.2% for description generation, and 12.4% for functional reasoning.

Industry Impact and Developer Recommendations

The release of the TRACE benchmark and FRM provides new research directions and tools for object understanding in post-fire environments. For developers, the plug-and-play nature of FRM makes it easy to integrate into existing systems, thereby improving model performance under complex degradation conditions. Additionally, the openness of the TRACE benchmark offers a standardized evaluation platform for researchers and engineers, fostering further advancements in the field.

---

Source: Hugging Face Daily Papers (2026-09-10)

— END —

Tags: #TRACE #Feature Recovery #Post-Fire Environment #Multimodal Vision-Language Models #FRM