segment-anything-model

Zero-shot image segmentation masks are generated by prompting the Segment Anything Model with points or bounding boxes to isolate arbitrary objects without dataset-specific fine-tuning. You provide a prompt, and the model returns candidate masks for immediate review and refinement.

Interactive annotation is supported by accepting point, bounding box, or previous mask prompts. This allows you to click or draw a box on an image, prompting the model to return candidate masks for fast semi-automated annotation workflows.

Automatic object masking is handled by the SamAutomaticMaskGenerator. It processes a static image to produce segmentation masks for all detected objects, eliminating the need for manual point or bounding box prompts during inference.

ONNX export is supported for deploying the Segment Anything Model in broader environments. You can choose from ViT-B, ViT-L, or ViT-H model variants and convert them to ONNX format outside of the standard PyTorch inference pipeline.

Images must be prepared in RGB format before running deterministic inference through SamPredictor or SamAutomaticMaskGenerator. This preprocessing step ensures the input matches the model's expected channel ordering for accurate zero-shot segmentation.