FastSAM is a CNN-based model designed for efficient image segmentation, offering a significantly faster alternative to the original Segment Anything Model (SAM). It targets researchers and developers needing high-speed segmentation with comparable performance, particularly for applications requiring real-time processing or large-scale dataset analysis.

How It Works

FastSAM leverages a CNN architecture, trained on a small fraction (2%) of the SA-1B dataset, to achieve its speed advantage. This approach allows for faster inference compared to SAM's transformer-based architecture, while maintaining competitive zero-shot transfer capabilities for tasks like edge detection and object proposal generation.

Quick Start & Requirements

• Install: Clone the repository, create a conda environment (conda create -n FastSAM python=3.9), activate it (conda activate FastSAM), and install dependencies (cd FastSAM; pip install -r requirements.txt). Install CLIP separately if using text prompts (pip install git+https://github.com/openai/CLIP.git).
• Prerequisites: Python >= 3.7, PyTorch >= 1.7, TorchVision >= 0.8. CUDA support for PyTorch/TorchVision is strongly recommended.
• Getting Started: Download model checkpoints. Run inference via python Inference.py with options for everything, text, box, or point prompts.
• Demos: HuggingFace Demo, Colab Demo, Replicate Demo & API.

Highlighted Details

• Achieves 50x higher run-time speed compared to SAM.
• Offers two model versions: FastSAM (YOLOv8x based) and FastSAM-s (YOLOv8s based).
• Supports various prompt modes: everything, text, box, and points.
• Demonstrates competitive zero-shot transfer performance on edge detection and object proposals.
• Integrates with Ultralytics (YOLOv8) Model Hub.

Maintenance & Community

The project has seen recent updates addressing edge jaggies and synchronizing with the Ultralytics project. It acknowledges contributions from various individuals and teams, including OpenXLab and Grounding-SAM. Links to demos and model zoo are provided.

Licensing & Compatibility

Licensed under the Apache 2.0 license, permitting commercial use and integration with closed-source projects.

Limitations & Caveats

While significantly faster, FastSAM's instance segmentation performance (AP) is lower than SAM and ViTDet-H on the COCO 2017 dataset. The project also notes an issue with edge jaggies, which has seen minor improvements.