Meta AI Releases Segment Anything Model 3 (SAM 3) for Promptable Concept Segmentation in Images and Videos

Meta AI Releases Segment Anything Model 3 (SAM 3) for Promptable Concept Segmentation in Images and Videos

Meta AI has released Segment Anything Model 3 (SAM 3), a new 848M parameter foundation model capable of promptable concept segmentation in both images and videos. Unlike previous models requiring manual pixel-level input, SAM 3 operates directly on visual concepts using text prompts, exemplars, points, or boxes.

Why This Matters

Traditional image segmentation relies on precise annotations, which are expensive and don’t scale to diverse real-world scenarios. SAM 3 addresses this by enabling segmentation based on concepts – like “red baseball cap” – across large datasets. The cost of manual annotation for large-scale video datasets can easily exceed millions of dollars; SAM 3 offers a path toward drastically reducing this expense.

Key Insights

• SA-Co Dataset: Meta introduced the SA-Co dataset with 270K unique concepts, 50x larger than previous benchmarks.
• Promptable Concept Segmentation (PCS): SAM 3 formalizes PCS, combining noun phrases with visual examples for accurate object identification.
• Presence Token: A novel component that improves precision by predicting whether a candidate mask corresponds to the requested concept, especially when dealing with similar entities.

Working Example

# Example usage (conceptual - actual API details would vary)
from segment_anything import sam3

# Load the model
model = sam3.Sam3(checkpoint="path/to/sam3_checkpoint.pth")

# Define a concept prompt
prompt = "yellow school bus"

# Provide an image
image = load_image("school_bus_image.jpg")

# Perform segmentation
masks, scores, logits = model.segment(image, prompt)

# Visualize the results
visualize_masks(image, masks, scores)

Practical Applications

• Autonomous Vehicles: Identifying pedestrians, traffic signs, and other objects in real-time using natural language prompts.
• Pitfall: Relying solely on text prompts can lead to ambiguity; incorporating exemplar images improves accuracy.

References:

• https://www.marktechpost.com/2025/11/20/meta-ai-releases-segment-anything-model-3-sam-3-for-promptable-concept-segmentation-in-images-and-videos/