It achieves comparable performance to the SAM method at 50 times faster.
(SAM) is a powerful vision-based model that can segment any object in an image based on user interaction cues. Once released, SAM gained huge traction in the computer vision community for its accuracy. However, SAM's extensive use of the computationally intensive Transformer (ViT) architecture limits its practical applications, especially in real-time scenarios.
FastSAM is an open-source image segmentation model trained on 2% of the SA-1B dataset trained by SAM. FastSAM reportedly runs 50 times faster than SAM