Abstract: Although the vision transformer-based methods (ViTs) exhibit an excellent performance than convolutional neural networks (CNNs) for image recognition tasks, their pixel-level semantic ...

Abstract: Recent contrastive multimodal vision-language models like CLIP have demonstrated robust open-world semantic understanding, becoming the standard image backbones for vision-language ...