Enhancing Vehicle Detection in Snowy Weather with Contrastive Learning

This thesis aims to improve vehicle detection in winter weather conditions by utilizingcontrastive learning with YOLO11n as the base detector. It explores various contrastivelearning frameworks, including SimCLR and DrLIM, and experiments with differentdefinitions of similar and dissimilar pairs to achieve effective representations. The thesisproposes a sideload-ICL-adaption framework by integrating a side CNN with YOLO11n.The side CNN is pretrained on a large amount of unannotated data using contrastivelearning. During the fine-tuning stage, additional features from the pretrained side CNNare fused with those from the YOLO backbone. Various fusion methods including staticaddition and dynamic gating are investigated. The proposed sideload-ICL-adaptionmodel outperforms the YOLO11n baseline by more than 3% in mAP50 on the testset. Additionally, freezing the YOLO11n backbone boosts mAP50 on the test set byapproximately 14%. The thesis also explores the learning behavior of each layer in theYOLO backbone to provide insights into fine-tuning.