| Summary: | <p>This thesis explores the frontier of visual perception in computer vision by leveraging the capabilities of Vision Transformers (ViTs) to create a unified framework that addresses cross-task and cross-granularity challenges. Drawing inspiration from the human visual system's ability to process visual information at varying levels of detail and the success of Transformers in Natural Language Processing (NLP), we aim to bridge the gap between broad visual concepts and their fine-grained counterparts. Our investigation is structured into three parts.</p>
<br>
<p>First, we delve into a range of training methods and architectures for ViTs, with the goal of gathering valuable insights. These insights are intended to guide the optimization of ViTs in the subsequent phase of our research, ensuring we build a strong foundation for enhancing their performance in complex visual tasks.</p>
<br>
<p>Second, our focus shifts towards the recognition of fine-grained visual concepts, employing precise annotations to delve deeper into the intricate details of visual scenes. Here, we tackle the challenge of discerning and classifying objects and pixels with remarkable accuracy, leveraging the foundational insights gained from our initial explorations of ViTs.</p>
<br>
<p>In the final part of our thesis, we demonstrate how language can serve as a bridge, enabling vision-language models, which are only trained to recognize images, to navigate countless visual concepts on fine-grained entities like objects and pixels without the need for fine-tuning.</p>
|
|---|