A Reddit user conducting a bachelor's thesis on fine-grained car classification found that the DINOv2 Giant model performed significantly worse than SigLIP2 SO400M when used as a frozen encoder for k-NN classification. Despite L2-normalized embeddings, DINOv2 yielded only 41% accuracy compared to SigLIP2's 92%. The user suspects DINOv2, trained via self-supervision, may require a trained head for fine-grained tasks, unlike contrastively trained models like SigLIP, and is seeking advice on its suitability for retrieval tasks. AI
IMPACT Highlights potential limitations of self-supervised models like DINOv2 for fine-grained retrieval tasks without further fine-tuning.
RANK_REASON User-reported benchmark comparison of different self-supervised and contrastive learning models for a specific downstream task. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →