Image-Based Vehicle Classification by Synergizing Features from Supervised and Self-Supervised Learning Paradigms

This paper introduces a novel approach to leverage features learned from both supervised and self-supervised paradigms, improve image classification tasks, specifically for vehicle classification. Two state-of-the-art learning methods, DINO data2vec, were evaluated compared their representation of images. The former contrasts local global views while the latter uses masked prediction on multi-layered representations. In case, is employed finetune pretrained YOLOR object detector detecting wheels, which definitive wheel positional are retrieved. representations these methods combined with task. Particularly, random masking strategy was utilized previously in harmony during training classifier. Our experiments show that data2vec-distilled representations, consistent our strategy, outperformed counterpart, resulting celebrated Top-1 accuracy 97.2% classifying 13 classes defined by Federal Highway Administration.