| Paper FR-EA-T32.4
Sivaraman, Anush Lakshman (Iowa State University), Adu-Gyamfi, Kojo (Iowa State University), Shihab, Ibne (Iowa State University), Sharma, Anuj (Iowa State Univeristy)
ClearVision: Leveraging CycleGAN and SigLIP-2 for Robust All-Weather Classification in Traffic Camera Imagery
Scheduled for presentation during the Regular Session "S32b-AI-Driven Traffic Monitoring, Safety, and Anomaly Detection" (FR-EA-T32), Friday, November 21, 2025,
14:30−14:50, Southport 2
2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia
This information is tentative and subject to change. Compiled on October 18, 2025
|
|
| Keywords AI, Machine Learning and Predictive Analytics for Traffic Incident Detection and Management, Deep Learning for Scene Understanding and Semantic Segmentation in Autonomous Vehicles, Real-world ITS Pilot Projects and Field Tests
Abstract
Adverse weather conditions pose significant challenges to safe transportation, highlighting the need for robust, real-time weather detection from traffic camera imagery. We present a novel framework that combines CycleGAN-based domain adaptation with efficient contrastive learning to enhance weather classification performance, particularly under low-light night-time conditions. Our method integrates a lightweight SigLIP-2 model—utilizing pairwise sigmoid loss to reduce computational overhead—with CycleGAN, which transforms night-time images into day-like representations while preserving critical weather cues. Evaluated on the Iowa Department of Transportation dataset, our approach demonstrates significant improvements over the EVA-02 baseline. While EVA-02 with CLIP achieves 96.55% per-class and day/night accuracy, it suffers from a notable day-night performance gap (97.21% vs. 63.40%). The integration of CycleGAN boosts EVA-02’s night-time accuracy to 82.45%. Our best-performing configuration—Vision-SigLIP-2 + Text-SigLIP-2 + CycleGAN + Contrastive Learning—achieves 85.90% night-time accuracy, 94.00% per-class accuracy, and 93.35% day/night accuracy. Moreover, it reduces training and inference times by 89% and 80%, respectively, compared to EVA-02, making it ideal for deployment in resource-constrained environments. By narrowing the day-night accuracy gap from 33.81 to 8.90 percentage points, our framework offers a scalable, efficient solution for all-weather classification using existing traffic camera infrastructure.
|
|
This site is protected by copyright and trademark laws under US and International
law.