Current Re-ID Model in DeepSORT Realtime

The default appearance feature extractor (Re-ID model) used in most DeepSORT implementations (e.g., deep-sort-realtime) is:

Mars-small128 (Market1501-trained)

• Architecture: A compact CNN (similar to ResNet but smaller).
• Input: Bounding box images (cropped detections).
• Output: 128-dimensional feature vector (embedding) for each object.
• Training Dataset: Market1501 (pedestrian Re-ID benchmark).
• Size: ~10MB (relatively heavy for mobile).

Why It’s Not Mobile-Friendly

1. Compute-Heavy: Designed for GPUs; slow on mobile CPUs.
2. Fixed Input Size: Typically 128x64 pixels, requiring resizing overhead.
3. No Hardware Optimization: Not quantized or pruned for edge devices.

Mobile-Friendly Re-ID Alternatives

Here are lighter models that can replace Mars-small128 for mobile deployment:

1. OSNet (Omni-Scale Network)

• Paper: OSNet (arXiv)

• Advantages:

  • Lightweight (1.8MB for osnet_x0_25).
  • Generalizes well to unseen objects (not just pedestrians).

• Implementation:

  # Install: pip install torchreid
  import torchreid
  model = torchreid.models.build_model("osnet_x0_25", pretrained=True)