Open-Vocabulary SAM: Segment and Recognize Twenty-thousand Classes Interactively
This work addresses the need for unified vision models that can handle both segmentation and recognition tasks interactively, representing an incremental improvement by integrating existing foundation models.
The paper tackles the problem of integrating segmentation and recognition in vision models by proposing Open-Vocabulary SAM, which combines SAM and CLIP to enable simultaneous interactive segmentation and recognition of around 22,000 classes, significantly outperforming naive baseline combinations.
The CLIP and Segment Anything Model (SAM) are remarkable vision foundation models (VFMs). SAM excels in segmentation tasks across diverse domains, whereas CLIP is renowned for its zero-shot recognition capabilities. This paper presents an in-depth exploration of integrating these two models into a unified framework. Specifically, we introduce the Open-Vocabulary SAM, a SAM-inspired model designed for simultaneous interactive segmentation and recognition, leveraging two unique knowledge transfer modules: SAM2CLIP and CLIP2SAM. The former adapts SAM's knowledge into the CLIP via distillation and learnable transformer adapters, while the latter transfers CLIP knowledge into SAM, enhancing its recognition capabilities. Extensive experiments on various datasets and detectors show the effectiveness of Open-Vocabulary SAM in both segmentation and recognition tasks, significantly outperforming the naïve baselines of simply combining SAM and CLIP. Furthermore, aided with image classification data training, our method can segment and recognize approximately 22,000 classes.