Benchmarking Foundation Models for Ocular Biometric Tasks : SVM and Few-Shot Classification using CLIP and DINOv2 Ocular Embeddings

Sammanfattning

This thesis investigated the use of frozen foundation-model image embeddings for ocular soft-biometric classification. The study focused on two tasks: binary gender classification and eight-class age-group classification from ocular/perioc- ular images. CLIP and DINOv2 were used as frozen image encoders, without fine- tuning, to extract fixed-dimensional visual embeddings. These embeddings were evaluated using two main downstream classification approaches: support vector machine-based classification and cosine-similarity-based classification. The SVM- based evaluation included LinearSVC trained with all available training data and a five-shot SVM setting, while the cosine-similarity-based evaluation included five- shot cosine similarity and mean-vector cosine similarity. The experiments were conducted across five predefined folds, and the five-shot experiments were re- peated over 10 runs to reduce the effect of random support-sample selection. The results showed that gender classification achieved higher accuracy than age- group classification across all evaluated methods. In the full-data SVM setting, LinearSVC achieved a mean gender accuracy of 75.15% using CLIP and 75.21% using DINOv2. For age-group classification, the corresponding mean accuracies were 42.32% for CLIP and 43.04% for DINOv2. The limited-data methods pro- duced lower results, while mean-vector cosine similarity performed between the five-shot methods and LinearSVC. The findings indicate that frozen foundation- model embeddings can provide useful representations for ocular gender classifi- cation, while age-group classification remains more challenging due to the higher number of classes, visual similarity between neighbouring age groups, and the limited information available in ocular/periocular crops