Janet H. Hsiao

Fine-grained CLIP fine-tuning with self-annotated region alignment

Chenyang Zhao, Wei Lin, Antoni B. Chan, Janet H. Hsiao

Contrastive Language-Image Pre-training (CLIP) has been shown to have limitations in its fine-grained dense feature representation, due to its pre-training focusing on matching the whole image to a text description. Considering the large data and computational burden in pre-train…Read more