HydroReasoner advances floodwater detection with knowledge-guided reasoning
Original title: Knowledge-Guided Floodwater Level Reasoning for Cloud-Edge Collaborative Monitoring
HydroReasoner uses Qwen2.5-VL and a floodwater knowledge graph to improve reasoning and generalize to varied scenes while running efficiently on edge devices.
Frequent urban floodwater threatens intelligent transportation systems. Existing vision methods lack explicit reasoning and struggle with viewpoint variations. HydroReasoner reformulates detection as visual reasoning and trains it with Group Relative Policy Optimization on a custom HydroKG knowledge graph. 4-bit quantization and cloud-edge routing cut memory and latency with minimal accuracy loss. The model achieves 4.0 and 13.6 percentage points higher accuracy than baselines on in-domain and out-of-domain tests.
- HydroReasoner builds on Qwen2.5-VL with KG path reward for fine-grained supervision
- 4-bit quantization reduces GPU memory by 57% and latency by 43%
- HydroLevel benchmark contains 3,000 floodwater images
- Performance drops only 4.4 percentage points out-of-domain
- Deterministic graph traversal supervises reasoning at fine granularity
From the original article
1 min read at arxiv.orgFrequent urban floodwater poses a serious threat to the safety and efficiency of intelligent transportation systems. Existing vision-based methods for floodwater level detection face two main limitations. First, the lack of explicit reasoning leads to poor detection performance and little interpretability.
By Zhengshu Zhou · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.