Meta has released the Segment Anything Model (SAM) version two, capable of segmenting any object in an image with the ability to generalize to objects not seen during training. The usage scenarios of SAM model include motion tracking, video effect creation, and background removal during meetings. The model requires inputs as points, boxes, or regions of the initial frame, both positive areas indicating desired objects and negative areas showing non-target objects. The model can automatically draw boundaries around all objects and remember what objects are being captured.
In addition to the model, Meta has also released the SA-V dataset for training SAM 2, containing over 600,000 object annotations across 51,000 videos from 47 countries, featuring diverse objects that may move in and out of frames. While SAM 2 demonstrates high efficiency, it may encounter limitations in tracking objects in crowded scenes, leading to occasional misidentifications.
The model is available for free usage under the Apache 2.0 license, while the SA-V dataset is open for use under the CC BY 4.0 license.
Source – AI at Meta
TLDR: Meta introduces SAM version two for object segmentation in images, along with the SA-V dataset, showcasing high performance but with limitations in crowded scenes.
Leave a Comment