Scalable and Efficient Multi-Weather Classification for Autonomous Driving with Coresets, Pruning, and Resolution Scaling

Published: 06 Mar 2025, Last Modified: 23 Apr 2025ICLR 2025 Workshop MLMP PosterEveryoneRevisionsBibTeXCC BY 4.0
Track: New scientific result
Keywords: Multi-weather classification, coreset selection, resolution scaling, pruning, model compression
Abstract:

Autonomous vehicles require robust perception systems capable of operating in diverse weather conditions, including snow, rain, fog, and storms. In this work, we present a scalable and efficient approach for multi-weather classification in autonomous driving, leveraging the WEDGE (WEather images by DALL-E GEneration) dataset. Our study investigates three complementary techniques to enhance classification performance and efficiency: Coreset Selection, Resolution Scaling, and Model Compression via Adaptive Pruning and Quantization. Specifically, we evaluate the impact of coreset selection methods (random and margin-based) at varying data fractions (e.g., 1, 0.75, 0.5, 0.25, 0.1), assess model robustness under low-resolution settings (224x224, 112x112, 56x56), and demonstrate that adaptive pruning combined with 8-bit quantization can reduce model size by up to 85% while maintaining competitive classification accuracy. Experimental results validate the effectiveness of our integrated approach, providing a scalable and robust solution for multi-weather classification. This work advances the feasibility of deploying perception models in real-world autonomous driving systems operating under adverse weather conditions and limited computational resources.

Submission Number: 25
Loading

OpenReview is a long-term project to advance science through improved peer review with legal nonprofit status. We gratefully acknowledge the support of the OpenReview Sponsors. © 2025 OpenReview