FloodReasonBench: Benchmarking VLM Reasoning Segmentation for Embodied Flood Response at the Edge
| Source: arXiv AI
Tags: VLM, reasoning segmentation, edge inference, disaster response, Jetson AGX Xavier
FloodReasonBench introduces a domain-specific benchmark for vision-language models doing pixel-level segmentation in real flood disaster scenarios on edge hardware, revealing sharp accuracy swings when model partitioning mismatches the deployment device.
Details
Existing VLM reasoning-segmentation benchmarks target generic visual scenes and ignore the constraints of disaster robotics: limited compute, split inference across network links, and domain-specific targets like submerged obstacles or stranded victims. FloodReasonBench fills that gap for flood-response platforms. The core contribution is FloodResponseSeg, a flood-specific dataset constructed from real-world scenes and response-relevant targets. The benchmark evaluates reasoning-segmentation pipelines under three operating conditions: lightweight visual encoding, hierarchical split inference, and compressed intermediate representations — all realistic for edge deployments. Testing on an NVIDIA Jetson AGX Xavier surfaces the accuracy-latency-energy-bandwidth tradeoff space. Generic models show wide partition-dependent accuracy variation, while flood-adapted models achieve a substantially tighter accuracy range across partitions. The paper is under review; code and dataset will be released only after acceptance.