SimonZeng7108/efficientsam3

What it solves

EfficientSAM3 addresses the high computational cost and large model size of the Segment Anything Model 3 (SAM3), making it more accessible for deployment on resource-constrained devices. It reduces the parameter count of both the vision and text encoders while maintaining competitive segmentation performance.

How it works

The project uses progressive hierarchical knowledge distillation to compress the heavy encoders of SAM3 into lightweight student models. It offers two main configurations:

1. EfficientSAM3 Full Models: Replaces both the vision encoder (using EfficientViT, RepViT, or TinyViT) and the text encoder (using MobileCLIP) to achieve up to a 90% reduction in size compared to ImageSAM3.
2. SAM3-LiteText: Keeps the original SAM3 vision encoder but replaces the heavy text encoder with a lightweight MobileCLIP variant, reducing the text encoder size by 88%.

Who it’s for

Researchers and developers who need high-quality image and video concept segmentation but require models that are smaller, faster, and more efficient for real-time or edge deployment.

Highlights

• Massive Compression: Full models are up to 90% smaller than ImageSAM3 (e.g., EV-M at 89.2M parameters).
• Flexible Architectures: Supports multiple lightweight backbones like RepViT, TinyViT, and EfficientViT.
• Distilled Text Encoders: Specifically optimized text encoders via MobileCLIP to maintain vision-language segmentation quality.
• Deployment Ready: Supports ONNX and TensorRT export for cross-platform deployment.