EdgeSAM - Efficient Segment Anything Model

EdgeSAM is an accelerated variant of the Segment Anything Model (SAM) optimized for edge devices using ONNX Runtime.

Model Files

• edge_sam_3x_encoder.onnx - Image encoder (1024x1024 input)
• edge_sam_3x_decoder.onnx - Mask decoder with prompt support

Usage

API Request Format

import requests
import base64

# Encode your image
with open("image.jpg", "rb") as f:
    image_b64 = base64.b64encode(f.read()).decode()

# Make request
response = requests.post(
    "https://YOUR-ENDPOINT-URL",
    json={
        "inputs": image_b64,
        "parameters": {
            "point_coords": [[512, 512]],  # Click point in 1024x1024 space
            "point_labels": [1],            # 1 = foreground, 0 = background
            "return_mask_image": True
        }
    }
)

result = response.json()

Response Format

[
  {
    "mask_shape": [1024, 1024],
    "has_object": true,
    "mask": "<base64_encoded_png>"
  }
]

Parameters

• point_coords: Array of [x, y] coordinates in 1024x1024 space (optional)
• point_labels: Array of labels (1=foreground, 0=background) corresponding to points (optional)
• box_coords: Bounding box [x1, y1, x2, y2] (optional, not yet implemented)
• return_mask_image: Return base64-encoded PNG mask (default: true)

Coordinate System

All coordinates should be in 1024x1024 space, regardless of original image size. The handler automatically resizes input images to 1024x1024 before processing.

Example: For a click at the center of any image, use [512, 512].

Local Testing

# Install dependencies
pip install -r requirements.txt

# Run test script
python test_handler.py

This will create:

• test_input.png - Test image with red circle
• test_output_mask.png - Generated segmentation mask
• test_output_overlay.png - Overlay visualization

Technical Details

• Input: RGB images (auto-resized to 1024x1024)
• Preprocessing: Normalized to [0, 1] range (/ 255.0)
• Hardware: Supports CUDA GPU with automatic CPU fallback
• Framework: ONNX Runtime Web compatible