The School of Computer Science is pleased to present…
Fine-Tuning Prototypes for Cross-Domain Few-Shot Image Classification Using Contrastive Objective
MSc Thesis Proposal by: Abhishek Mahajan
Date: Monday, May 26th, 2025
Time: 9:00 AM
Location: Essex Hall, Room 122
Cross-domain few-shot learning (CFC) seeks to enable accurate classification in novel visual domains using only a few labelled samples per class. However, it remains a challenging task due to prototype misalignment and domain-induced embedding distortions, in addition to a lack of purified data. While the state-of-the-art transformation network effectively realigns prototypes in one-shot settings, it is not designed for broader domain generalization or transformer-based architectures in cross-domain settings. We introduce CosRestViT, a prototype-aware few-shot learning framework designed to tackle the problems of scarce data with a lot of outliers. Our method integrates a custom transformation network with a Vision Transformer (ViT) backbone to recalibrate noisy embeddings and align them with class-level prototypes. To enhance semantic consistency in the embedding space, we fine-tune the model using a hybrid contrastive objective that combines cross-entropy loss with CoSENT’s ranking loss. We pretrain CosRestViT on the large-scale base dataset using a prototype-aware contrastive loss to enhance the representational power of ViT and ensure effective clustering of semantically similar samples. The model is then evaluated on the Meta-Dataset benchmark under two training regimes: (i) training on all seen datasets, and (ii) training exclusively on ImageNet for transfer learning. Experimental results demonstrate that CosRestViT outperforms existing baselines across both seen and unseen domains, achieving superior generalization in low-data scenarios.
Internal Reader: Dr. Dan Wu
External Reader: Dr. Abdul A. Hussein
Advisor: Dr. Ziad Kobti
