Open Internet by MindsNet

Improving Fine-Grained Classification with DINOv2

The author is experiencing a significant performance gap between SigLIP and DINOv2 in fine-grained car classification tasks, with DINOv2 achieving only 41% accuracy. The author is seeking advice on how to improve DINOv2's performance, particularly with regards to using a linear probe or adjusting the embedding layer.

Computing & Technology, Computer Science, Machine Learning