Smaller Self-Supervised ViTs Localize Better than Larger Ones

Arxiv CS.CV (computer vision)

A new arXiv study finds that smaller self-supervised Vision Transformers produce better foreground object localization than larger models. The paper attributes this to differences in attention map dynamics during training.

vision-transformerself-supervised

Arxiv CS.CV (computer vision)·Sreehari RammohanHuy HaCarl Vondrick

8 days ago