How to Integrate a Mobile-Friendly Re-ID Model with DeepSORT

If your objects are similar to the model’s training data (e.g., pedestrians/vehicles), you can directly use pre-trained models like:

• OSNet (general-purpose Re-ID)
• MobileNetV3 (trained on Market1501 or VeRi-776)
• FastReID (pre-trained on person/vehicle datasets)

Steps:

1. Install Dependencies:

   pip install deep-sort-realtime torchreid  # For OSNet
   

2. Replace the Embedder:

   from deep_sort_realtime.embedder import CustomEmbedder
   import torchreid
   
   class MobileReIDEmbedder(CustomEmbedder):
       def __init__(self):
           self.model = torchreid.models.build_model("osnet_x0_25", pretrained=True)
           self.model.eval()  # Disable dropout/BatchNorm
   
       def embed(self, crops):
           # crops: List of RGB numpy arrays (bounding box images)
           # Preprocess: Resize to 128x64, normalize, convert to tensor
           embeddings = self.model(crops)  # Returns [N, 128] features
           return embeddings.cpu().numpy()
   
   # Initialize DeepSORT with the custom embedder
   tracker = DeepSort(embedder=MobileReIDEmbedder())
   

Optimization Tips for Mobile

1. Quantize the Model:

   # Convert to INT8 (TensorFlow Lite/PyTorch Quantization)
   torch.quantization.quantize_dynamic(model, {torch.nn.Linear}, dtype=torch.qint8)
   

2. Use Hardware Acceleration:

   • CoreML (iOS) / NNAPI (Android) / ONNX Runtime.

3. Reduce Input Size:

   • Resize crops to 64x32 pixels for faster inference.