Thanks for sharing! The work unveils how SAM, devoid of semantic labels during training, grasps high-level semantics for captioning. Thus, we can efficiently au

Thanks for sharing! The work unveils how SAM, devoid of semantic labels during training, grasps high-level semantics for captioning. Thus, we can efficiently augment SAM for regional captioning. Mark a finding towards understanding vision capabilities only from low-level data! For details, please refer to the project page: https://xk-huang.github.io/segment-caption-anything/ https://xk-huang.github.io/segment-caption-anything/