Segment Anything (SAM)

Segment Anything uses a vision transformer encoder-decoder architecture that accepts flexible prompts (points, bounding boxes, text, or masks) to segment any object in an image without task-specific fine-tuning. The model encodes the image once with a ViT backbone, then uses a lightweight mask decoder that processes prompt embeddings to generate segmentation masks in real-time. This prompt-based approach enables zero-shot segmentation across diverse object categories without retraining.

SAM includes an automatic mask generation mode that systematically grids the image with point prompts and runs the segmentation decoder on each grid cell to produce a comprehensive set of non-overlapping masks covering all salient objects. The system uses non-maximum suppression and confidence filtering to deduplicate overlapping masks and retain only high-quality segmentations. This enables one-shot full-image instance segmentation without manual prompting.

SAM uses a Vision Transformer (ViT) backbone to encode images into dense feature maps that capture multi-scale visual information. The encoder processes the full image at once, producing hierarchical feature representations that preserve spatial structure while enabling the lightweight decoder to generate masks from arbitrary prompts. This design choice enables efficient amortization of computation across multiple prompts on the same image.

SAM's mask decoder is a small transformer-based module that fuses image features from the ViT encoder with prompt embeddings (points, boxes, or masks) to generate segmentation masks. The decoder uses cross-attention mechanisms to align prompt information with image features, producing binary masks and confidence scores in real-time. This lightweight design enables fast inference and enables the decoder to be trained independently from the frozen image encoder.

SAM's decoder can generate multiple mask candidates for ambiguous prompts (e.g., a point on an object boundary could belong to multiple objects). The model produces a primary mask plus one or more alternative masks with associated confidence scores, enabling downstream systems to rank or select the most appropriate segmentation. This design acknowledges that segmentation is inherently ambiguous and provides tools for disambiguation.

SAM was trained on SA-1B, a dataset of 1.1 billion segmentation masks automatically generated from 11 million images using an iterative process: initial SAM predictions were refined with human feedback, then used to generate additional masks via automatic prompting. This dataset construction process demonstrates how to bootstrap large-scale segmentation annotations without manual labeling, enabling SAM's zero-shot generalization across diverse object categories and image domains.

SAM achieves zero-shot generalization across diverse image domains (natural images, medical imaging, satellite imagery, etc.) by leveraging a ViT encoder pre-trained on large-scale vision datasets. The encoder learns domain-agnostic visual features that transfer effectively to new domains without fine-tuning, while the lightweight mask decoder is trained on diverse segmentation masks from SA-1B. This design enables SAM to segment objects in domains not seen during training.

SAM can be fine-tuned on domain-specific segmentation data by training the lightweight mask decoder on labeled masks from the target domain while keeping the ViT encoder frozen. This approach enables rapid adaptation to specialized domains (medical imaging, satellite imagery, etc.) with limited labeled data, reducing fine-tuning time and data requirements compared to training end-to-end models. The frozen encoder preserves domain-agnostic visual features while the decoder learns domain-specific segmentation patterns.

SAM supports interactive refinement workflows where users provide initial prompts (points or boxes), review the generated masks, and iteratively refine prompts to correct segmentation errors. The system reuses the frozen image encoding across refinement iterations, enabling sub-100ms mask generation for each refinement step. This design enables efficient human-in-the-loop annotation where users guide the model toward correct segmentations through iterative feedback.

SAM supports various inference optimizations including model quantization (INT8, FP16), knowledge distillation to smaller models, and hardware-specific optimizations (ONNX, TensorRT) to enable deployment on resource-constrained devices. These optimizations reduce model size by 4-8x and inference latency by 2-4x while maintaining segmentation quality, enabling SAM deployment on mobile devices, edge hardware, and real-time applications. The frozen encoder design facilitates efficient optimization by decoupling image encoding from mask generation.