segment-anything-onnx 0.0.5

Run Segment Anything with ONNX models

segment-anything-onnx

Use ONNX models for Segment Anything inference.

Special thanks to:

• Viet-Anh NGUYEN for creating the SAM Exporter libraries from which these files are derived.
• The Segment Anything team at Meta AI Research

Usage from Source

1. Clone Segment Anything ONNX from Github

   git clone [e-mail]:whatstyle-ai/segment-anything-onnx.git
   cd segment-anything-onnx
   pip install -e .
   

2. Use the SAM Exporter to generate the ONNX models
3. Copy the ONNX models to the segment-anything-onnx/models directory
4. Predict some masks

   cd segment-anything-onnx
   ./demo.sh
   

Usage from pip install

1. Use the SAM Exporter to generate the ONNX models, or obtain the ONNX models from another source
2. Copy the ONNX models to a "models" directory, such as:

   models/sam_vit_l_0b3195.encoder.onnx
   models/sam_vit_l_0b3195.decoder.onnx
   

3. Install Segment Anything ONNX using pip:

   pip install segment-anything-onnx
   

4. Predict a mask:

   import cv2
   import urllib.request
   import numpy as np
   
   from segment_anything_onnx.inference import predict_masks
   
   
   def load_image(uri):
       if( uri.startswith('https://') or uri.startswith('http://') ):
           req = urllib.request.urlopen(uri)
           arr = np.asarray(bytearray(req.read()), dtype=np.uint8)
           img = cv2.imdecode(arr, -1) # 'Load it as it is'
           return img
       else: 
           return cv2.imread(uri)  # uri is just a local file path
   
   
   encoder_model_path = './models/sam_vit_l_0b3195.encoder.onnx'
   decoder_model_path = './models/sam_vit_l_0b3195.decoder.onnx'
   image = load_image( 'https://raw.githubusercontent.com/whatstyle-ai/segment-anything-onnx/main/examples/laura.jpg' )
   prompt = [
       { 'type': 'point', 'data': [1750, 300], 'label': 0 },
       { 'type': 'rectangle', 'data': [611, 655, 2712, 4500] }
   ]
   options = {
       'show': True,
       'output': './output/laura-L.png'
   }
   
   predict_masks( encoder_model_path, decoder_model_path, image, prompt, options )