---
library_name: pytorch
license: mit
tags:
- backbone
- real_time
- android
pipeline_tag: video-object-tracking

---

![](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/osnet/web-assets/model_demo.png)

# OSNet: Optimized for Qualcomm Devices

OSNet is a compact convolutional architecture for person re-identification. It extracts discriminative embeddings that can be compared with cosine similarity for person matching across camera views.

This is based on the implementation of OSNet found [here](https://github.com/KaiyangZhou/deep-person-reid).
This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/v0.61.0/src/qai_hub_models/models/osnet) library to export with custom configurations. More details on model performance across various devices, can be found [here](#performance-summary).

Qualcomm AI Hub Models uses [Qualcomm AI Hub Workbench](https://workbench.aihub.qualcomm.com) to compile, profile, and evaluate this model. [Sign up](https://myaccount.qualcomm.com/signup) to run these models on a hosted Qualcomm® device.

## Getting Started
There are two ways to deploy this model on your device:

### Option 1: Download Pre-Exported Models

Below are pre-exported model assets ready for deployment.

| Runtime | Precision | Chipset | SDK Versions | Download |
|---|---|---|---|---|
| ONNX | float | Universal | QAIRT 2.45, ONNX Runtime 1.27.1 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/osnet/releases/v0.61.0/osnet-onnx-float.zip)
| QNN_DLC | float | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/osnet/releases/v0.61.0/osnet-qnn_dlc-float.zip)
| QNN_DLC | w8a8 | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/osnet/releases/v0.61.0/osnet-qnn_dlc-w8a8.zip)
| TFLITE | float | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/osnet/releases/v0.61.0/osnet-tflite-float.zip)
| TFLITE | w8a8 | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/osnet/releases/v0.61.0/osnet-tflite-w8a8.zip)

For more device-specific assets and performance metrics, visit **[OSNet on Qualcomm® AI Hub](https://aihub.qualcomm.com/models/osnet)**.


### Option 2: Export with Custom Configurations

Use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/v0.61.0/src/qai_hub_models/models/osnet) Python library to compile and export the model with your own:
- Custom weights (e.g., fine-tuned checkpoints)
- Custom input shapes
- Target device and runtime configurations

This option is ideal if you need to customize the model beyond the default configuration provided here.

See our repository for [OSNet on GitHub](https://github.com/qualcomm/ai-hub-models/blob/v0.61.0/src/qai_hub_models/models/osnet) for usage instructions.

## Model Details

**Model Type:** Model_use_case.video_object_tracking

**Model Stats:**
- Input resolution: 256x128
- Model checkpoint: osnet_x1_0_market_256x128_amsgrad_ep150_stp60_lr0.0015_b64_fb10_softmax_labelsmooth_flip.pth
- Model size (float): 8.69 MB
- Model size (w8a8): 3.57 MB
- Number of parameters: 2.68M

## Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit
|---|---|---|---|---|---|---
| OSNet | ONNX | float | Snapdragon® X2 Elite | 0.497 ms | 1 - 1 MB | NPU
| OSNet | ONNX | float | Snapdragon® X Elite | 0.934 ms | 0 - 0 MB | NPU
| OSNet | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 0.617 ms | 0 - 71 MB | NPU
| OSNet | ONNX | float | Snapdragon® 8 Gen 1 Mobile | 1.598 ms | 0 - 68 MB | NPU
| OSNet | ONNX | float | Qualcomm® Dragonwing™ IQ-8275 | 1.406 ms | 0 - 4 MB | NPU
| OSNet | ONNX | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 0.89 ms | 0 - 95 MB | NPU
| OSNet | ONNX | float | Qualcomm® QCS8450 | 1.598 ms | 0 - 68 MB | NPU
| OSNet | ONNX | float | Qualcomm® Dragonwing™ IQ-9075 | 1.395 ms | 0 - 3 MB | NPU
| OSNet | ONNX | float | Qualcomm® Dragonwing™ IQ-X7181 | 0.934 ms | 0 - 0 MB | NPU
| OSNet | ONNX | float | Qualcomm® Dragonwing™ Q-8750 | 0.49 ms | 0 - 53 MB | NPU
| OSNet | ONNX | float | Snapdragon® 8 Elite Mobile | 0.49 ms | 0 - 53 MB | NPU
| OSNet | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.456 ms | 0 - 54 MB | NPU
| OSNet | QNN_DLC | float | Snapdragon® X2 Elite | 0.643 ms | 0 - 0 MB | NPU
| OSNet | QNN_DLC | float | Snapdragon® X Elite | 1.243 ms | 0 - 0 MB | NPU
| OSNet | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 0.705 ms | 0 - 58 MB | NPU
| OSNet | QNN_DLC | float | Snapdragon® 8 Gen 1 Mobile | 2.022 ms | 0 - 62 MB | NPU
| OSNet | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-8275 | 1.441 ms | 0 - 3 MB | NPU
| OSNet | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-8275 | 4.564 ms | 0 - 43 MB | NPU
| OSNet | QNN_DLC | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 1.052 ms | 0 - 10 MB | NPU
| OSNet | QNN_DLC | float | Qualcomm® SA8775P | 1.68 ms | 0 - 47 MB | NPU
| OSNet | QNN_DLC | float | Qualcomm® SA8650P | 1.68 ms | 0 - 47 MB | NPU
| OSNet | QNN_DLC | float | Qualcomm® SA8255P | 1.68 ms | 0 - 47 MB | NPU
| OSNet | QNN_DLC | float | Qualcomm® QCS8450 | 2.022 ms | 0 - 62 MB | NPU
| OSNet | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-9075 | 1.482 ms | 0 - 2 MB | NPU
| OSNet | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-X7181 | 1.243 ms | 0 - 0 MB | NPU
| OSNet | QNN_DLC | float | Qualcomm® Dragonwing™ Q-8750 | 0.533 ms | 0 - 49 MB | NPU
| OSNet | QNN_DLC | float | Qualcomm® SA7255P | 4.564 ms | 0 - 43 MB | NPU
| OSNet | QNN_DLC | float | Qualcomm® SA8295P | 2.187 ms | 0 - 40 MB | NPU
| OSNet | QNN_DLC | float | Snapdragon® 8 Elite Mobile | 0.533 ms | 0 - 49 MB | NPU
| OSNet | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.428 ms | 0 - 47 MB | NPU
| OSNet | QNN_DLC | w8a8 | Snapdragon® X2 Elite | 0.351 ms | 0 - 0 MB | NPU
| OSNet | QNN_DLC | w8a8 | Snapdragon® X Elite | 0.662 ms | 0 - 0 MB | NPU
| OSNet | QNN_DLC | w8a8 | Snapdragon® 8 Gen 3 Mobile | 0.371 ms | 0 - 56 MB | NPU
| OSNet | QNN_DLC | w8a8 | Snapdragon® 8 Gen 1 Mobile | 0.671 ms | 0 - 62 MB | NPU
| OSNet | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ QCS6490 | 1.494 ms | 0 - 2 MB | NPU
| OSNet | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ IQ-8275 | 0.548 ms | 0 - 3 MB | NPU
| OSNet | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ IQ-8275 | 1.264 ms | 0 - 44 MB | NPU
| OSNet | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 0.533 ms | 0 - 2 MB | NPU
| OSNet | QNN_DLC | w8a8 | Qualcomm® SA8775P | 0.747 ms | 0 - 48 MB | NPU
| OSNet | QNN_DLC | w8a8 | Qualcomm® SA8650P | 0.747 ms | 0 - 48 MB | NPU
| OSNet | QNN_DLC | w8a8 | Qualcomm® SA8255P | 0.747 ms | 0 - 48 MB | NPU
| OSNet | QNN_DLC | w8a8 | Qualcomm® QCS8450 | 0.671 ms | 0 - 62 MB | NPU
| OSNet | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ IQ-9075 | 0.62 ms | 0 - 2 MB | NPU
| OSNet | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ IQ-X7181 | 0.662 ms | 0 - 0 MB | NPU
| OSNet | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ Q-6690 | 2.803 ms | 0 - 41 MB | NPU
| OSNet | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ Q-7790 | 0.538 ms | 0 - 44 MB | NPU
| OSNet | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ Q-8750 | 0.282 ms | 0 - 43 MB | NPU
| OSNet | QNN_DLC | w8a8 | Qualcomm® SA7255P | 1.264 ms | 0 - 44 MB | NPU
| OSNet | QNN_DLC | w8a8 | Qualcomm® SA8295P | 1.016 ms | 0 - 43 MB | NPU
| OSNet | QNN_DLC | w8a8 | Snapdragon® 8 Elite Mobile | 0.282 ms | 0 - 43 MB | NPU
| OSNet | QNN_DLC | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.23 ms | 0 - 45 MB | NPU
| OSNet | QNN_DLC | w8a8 | Snapdragon® 7 Gen 4 Mobile | 0.538 ms | 0 - 44 MB | NPU
| OSNet | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 0.722 ms | 0 - 59 MB | NPU
| OSNet | TFLITE | float | Snapdragon® 8 Gen 1 Mobile | 2.034 ms | 0 - 63 MB | NPU
| OSNet | TFLITE | float | Qualcomm® Dragonwing™ IQ-8275 | 1.468 ms | 0 - 8 MB | NPU
| OSNet | TFLITE | float | Qualcomm® Dragonwing™ IQ-8275 | 4.587 ms | 0 - 45 MB | NPU
| OSNet | TFLITE | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 1.105 ms | 0 - 2 MB | NPU
| OSNet | TFLITE | float | Qualcomm® SA8775P | 1.757 ms | 0 - 48 MB | NPU
| OSNet | TFLITE | float | Qualcomm® SA8650P | 1.757 ms | 0 - 48 MB | NPU
| OSNet | TFLITE | float | Qualcomm® SA8255P | 1.757 ms | 0 - 48 MB | NPU
| OSNet | TFLITE | float | Qualcomm® QCS8450 | 2.034 ms | 0 - 63 MB | NPU
| OSNet | TFLITE | float | Qualcomm® Dragonwing™ IQ-9075 | 1.57 ms | 0 - 7 MB | NPU
| OSNet | TFLITE | float | Qualcomm® Dragonwing™ Q-8750 | 0.537 ms | 0 - 44 MB | NPU
| OSNet | TFLITE | float | Qualcomm® SA7255P | 4.587 ms | 0 - 45 MB | NPU
| OSNet | TFLITE | float | Qualcomm® SA8295P | 2.263 ms | 0 - 39 MB | NPU
| OSNet | TFLITE | float | Snapdragon® 8 Elite Mobile | 0.537 ms | 0 - 44 MB | NPU
| OSNet | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.443 ms | 0 - 45 MB | NPU
| OSNet | TFLITE | w8a8 | Snapdragon® 8 Gen 3 Mobile | 0.315 ms | 0 - 52 MB | NPU
| OSNet | TFLITE | w8a8 | Snapdragon® 8 Gen 1 Mobile | 0.588 ms | 0 - 59 MB | NPU
| OSNet | TFLITE | w8a8 | Qualcomm® Dragonwing™ QCS6490 | 2.429 ms | 0 - 12 MB | NPU
| OSNet | TFLITE | w8a8 | Qualcomm® Dragonwing™ IQ-8275 | 0.492 ms | 0 - 5 MB | NPU
| OSNet | TFLITE | w8a8 | Qualcomm® Dragonwing™ IQ-8275 | 1.142 ms | 0 - 41 MB | NPU
| OSNet | TFLITE | w8a8 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 0.442 ms | 0 - 73 MB | NPU
| OSNet | TFLITE | w8a8 | Qualcomm® SA8775P | 0.691 ms | 0 - 44 MB | NPU
| OSNet | TFLITE | w8a8 | Qualcomm® SA8650P | 0.691 ms | 0 - 44 MB | NPU
| OSNet | TFLITE | w8a8 | Qualcomm® SA8255P | 0.691 ms | 0 - 44 MB | NPU
| OSNet | TFLITE | w8a8 | Qualcomm® QCS8450 | 0.588 ms | 0 - 59 MB | NPU
| OSNet | TFLITE | w8a8 | Qualcomm® Dragonwing™ IQ-9075 | 0.551 ms | 0 - 5 MB | NPU
| OSNet | TFLITE | w8a8 | Qualcomm® Dragonwing™ Q-6690 | 3.546 ms | 0 - 40 MB | NPU
| OSNet | TFLITE | w8a8 | Qualcomm® Dragonwing™ Q-7790 | 0.914 ms | 0 - 42 MB | NPU
| OSNet | TFLITE | w8a8 | Qualcomm® Dragonwing™ Q-8750 | 0.257 ms | 0 - 40 MB | NPU
| OSNet | TFLITE | w8a8 | Qualcomm® SA7255P | 1.142 ms | 0 - 41 MB | NPU
| OSNet | TFLITE | w8a8 | Qualcomm® SA8295P | 0.922 ms | 0 - 40 MB | NPU
| OSNet | TFLITE | w8a8 | Snapdragon® 8 Elite Mobile | 0.257 ms | 0 - 40 MB | NPU
| OSNet | TFLITE | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.215 ms | 0 - 43 MB | NPU
| OSNet | TFLITE | w8a8 | Snapdragon® 7 Gen 4 Mobile | 0.914 ms | 0 - 42 MB | NPU

## License
* The license for the original implementation of OSNet can be found
  [here](https://github.com/KaiyangZhou/deep-person-reid/blob/master/LICENSE).

## References
* [Omni-Scale Feature Learning for Person Re-Identification](https://arxiv.org/abs/1905.00953)
* [Source Model Implementation](https://github.com/KaiyangZhou/deep-person-reid)

## Community
* Join [our AI Hub Slack community](https://aihub.qualcomm.com/community/slack) to collaborate, post questions and learn more about on-device AI.
* For questions or feedback please [reach out to us](mailto:ai-hub-support@qti.qualcomm.com).