🌟 Summary

This update introduces flexible image size inference for SAM2 using the ultralytics package, alongside improvements to testing and documentation.

📊 Key Changes

• Flexible SAM2 Image Size Inference: SAM2 now supports custom image sizes through the ultralytics package, allowing inference at sizes like 640x640 instead of the default 1024x1024.
• Testing Enhancement: Updated CI testing workflow for version-specific compatibility.
• Documentation Update: Refreshed documentation with higher quality images.

🎯 Purpose & Impact

• Enhanced Flexibility: Users can now run SAM2 inference using the ultralytics package with custom imgsz values (e.g., imgsz=640), offering significant advantages:
  • Faster processing times for smaller images
  • Reduced memory usage, enabling inference on devices with limited resources
  • Maintains good segmentation quality while allowing size-performance tradeoffs
• Improved Efficiency: The ability to use smaller image sizes (like 640x640) instead of the default 1024x1024 can lead to faster inference without significant loss in accuracy for many use cases.
• Broader Accessibility: This update makes SAM2 more accessible for users with varying computational resources, as they can now adjust the image size based on their specific needs and hardware constraints.

💻 Usage Example

from ultralytics import SAM

# Initialize SAM model
model = SAM('sam2_b.pt')

# Run inference with custom image size
results = model('path/to/image.jpg', imgsz=640)

This update significantly enhances SAM2's versatility within the ultralytics ecosystem, allowing users to fine-tune the balance between speed and accuracy based on their specific requirements.

What's Changed

• Add retry step to failed Conda tests by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/15884
• Use AVIF banner images by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/15876
• Remove image "?" args by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/15891
• fix HUB download and train by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/15896
• Optimize Docs images by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/15900
• Run Conda tests with aligned tag/version by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/15904
• Adding missing datasets information to docs by @jk4e in https://github.com/ultralytics/ultralytics/pull/15916
• ultralytics 8.2.84 new SAM flexible imgsz inference by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/15882

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.2.83...v8.2.84