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: Aligned text-image encoders such as CLIP have become the de-facto model for vision-language tasks. Further-more, modality-specific encoders achieve impressive per-formances in their ...