SAM (Segment Anything Model)
A foundation model by Meta for universal image segmentation that can segment any object in an image with zero-shot capability.
SAM (Segment Anything) is Meta's foundation model for universal segmentation – segments any object zero-shot via click, box, or text prompt.
Explanation
SAM was trained on over 1 billion masks and can be prompted via points, boxes, or text. SAM 2 (2024) extends this to video.
Marketing Relevance
SAM democratizes segmentation – no domain-specific training data needed. Useful for creative tools, medical imaging, and data annotation.
Example
A design tool uses SAM to cut out objects in photos – the user simply clicks on the desired object.
Common Pitfalls
High compute for real-time. Weaknesses with abstract/textureless regions. Finest edges sometimes imprecise.
Origin & History
Meta released SAM in April 2023 with the SA-1B dataset (1 billion masks). It was the first "foundation model" for segmentation. SAM 2 (July 2024) extended to video segmentation with memory and tracking.
Comparisons & Differences
SAM (Segment Anything Model) vs. U-Net
U-Net needs domain-specific training. SAM is a foundation model and works zero-shot on any image.
SAM (Segment Anything Model) vs. Mask R-CNN
Mask R-CNN detects and segments predefined classes. SAM segments any object class-agnostically.