---
library_name: pytorch
license: other
tags:
- real_time
- android
pipeline_tag: keypoint-detection

---

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

# SixDRepNet: Optimized for Qualcomm Devices

6DRepNet predicts head pose (pitch, yaw, roll) from a face image using a RepVGG-B1g2 backbone and a continuous 6D rotation representation, achieving robust and accurate head pose estimation.

This is based on the implementation of SixDRepNet found [here](https://github.com/thohemp/6DRepNet).
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/v0.57.1/src/qai_hub_models/models/sixd_repnet) 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 1.25.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/sixd_repnet/releases/v0.57.1/sixd_repnet-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/sixd_repnet/releases/v0.57.1/sixd_repnet-qnn_dlc-float.zip)
| TFLITE | float | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/sixd_repnet/releases/v0.57.1/sixd_repnet-tflite-float.zip)

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


### Option 2: Export with Custom Configurations

Use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/v0.57.1/src/qai_hub_models/models/sixd_repnet) 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 [SixDRepNet on GitHub](https://github.com/qualcomm/ai-hub-models/blob/v0.57.1/src/qai_hub_models/models/sixd_repnet) for usage instructions.

## Model Details

**Model Type:** Model_use_case.pose_estimation

**Model Stats:**
- Input resolution: 224x224
- Number of parameters: 15.3M
- Model size (float): 58.4 MB

## Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit
|---|---|---|---|---|---|---
| face_detector | ONNX | float | Snapdragon® X2 Elite | 1.914 ms | 208 - 208 MB | NPU
| face_detector | ONNX | float | Snapdragon® X Elite | 3.7 ms | 177 - 177 MB | NPU
| face_detector | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 2.188 ms | 2 - 156 MB | NPU
| face_detector | ONNX | float | Snapdragon® 8 Gen 1 Mobile | 6.832 ms | 1 - 165 MB | NPU
| face_detector | ONNX | float | Qualcomm® QCS8550 (Proxy) | 3.519 ms | 0 - 15 MB | NPU
| face_detector | ONNX | float | Qualcomm® QCS8450 | 6.832 ms | 1 - 165 MB | NPU
| face_detector | ONNX | float | Snapdragon® 8 Elite Mobile | 1.722 ms | 0 - 166 MB | NPU
| face_detector | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 1.817 ms | 0 - 159 MB | NPU
| face_detector | ONNX | float | Qualcomm® QCS9075 | 5.43 ms | 5 - 50 MB | NPU
| face_detector | ONNX | float | Qualcomm® QCS8750 | 1.722 ms | 0 - 166 MB | NPU
| face_detector | ONNX | float | Qualcomm® QCS7181 | 3.7 ms | 177 - 177 MB | NPU
| face_detector | QNN_DLC | float | Snapdragon® X2 Elite | 5.851 ms | 5 - 5 MB | NPU
| face_detector | QNN_DLC | float | Snapdragon® X Elite | 16.58 ms | 5 - 5 MB | NPU
| face_detector | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 9.334 ms | 5 - 177 MB | NPU
| face_detector | QNN_DLC | float | Snapdragon® 8 Gen 1 Mobile | 24.527 ms | 3 - 174 MB | NPU
| face_detector | QNN_DLC | float | Qualcomm® QCS8275 | 28.241 ms | 0 - 150 MB | NPU
| face_detector | QNN_DLC | float | Qualcomm® QCS8550 (Proxy) | 15.615 ms | 5 - 6 MB | NPU
| face_detector | QNN_DLC | float | Qualcomm® QCS8450 | 24.527 ms | 3 - 174 MB | NPU
| face_detector | QNN_DLC | float | Snapdragon® 8 Elite Mobile | 6.918 ms | 0 - 154 MB | NPU
| face_detector | QNN_DLC | float | Qualcomm® SA8295P | 20.645 ms | 0 - 150 MB | NPU
| face_detector | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 5.455 ms | 5 - 158 MB | NPU
| face_detector | QNN_DLC | float | Qualcomm® SA7255P | 28.241 ms | 0 - 150 MB | NPU
| face_detector | QNN_DLC | float | Qualcomm® QCS9075 | 19.595 ms | 5 - 12 MB | NPU
| face_detector | QNN_DLC | float | Qualcomm® QCS8750 | 6.918 ms | 0 - 154 MB | NPU
| face_detector | QNN_DLC | float | Qualcomm® QCS7181 | 16.58 ms | 5 - 5 MB | NPU
| face_detector | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 9.267 ms | 1 - 168 MB | NPU
| face_detector | TFLITE | float | Snapdragon® 8 Gen 1 Mobile | 24.64 ms | 0 - 172 MB | NPU
| face_detector | TFLITE | float | Qualcomm® QCS8275 | 28.279 ms | 1 - 150 MB | NPU
| face_detector | TFLITE | float | Qualcomm® QCS8550 (Proxy) | 15.655 ms | 1 - 3 MB | NPU
| face_detector | TFLITE | float | Qualcomm® SA8775P | 7.778 ms | 1 - 32 MB | GPU
| face_detector | TFLITE | float | Qualcomm® SA8650P | 7.778 ms | 1 - 32 MB | GPU
| face_detector | TFLITE | float | Qualcomm® SA8255P | 7.778 ms | 1 - 32 MB | GPU
| face_detector | TFLITE | float | Qualcomm® QCS8450 | 24.64 ms | 0 - 172 MB | NPU
| face_detector | TFLITE | float | Snapdragon® 8 Elite Mobile | 6.919 ms | 1 - 150 MB | NPU
| face_detector | TFLITE | float | Qualcomm® SA8295P | 20.674 ms | 1 - 151 MB | NPU
| face_detector | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 5.488 ms | 1 - 154 MB | NPU
| face_detector | TFLITE | float | Qualcomm® SA7255P | 28.279 ms | 1 - 150 MB | NPU
| face_detector | TFLITE | float | Qualcomm® QCS9075 | 19.262 ms | 0 - 9 MB | NPU
| face_detector | TFLITE | float | Qualcomm® QCS8750 | 6.919 ms | 1 - 150 MB | NPU
| pose_estimator | ONNX | float | Snapdragon® X2 Elite | 1.299 ms | 212 - 212 MB | NPU
| pose_estimator | ONNX | float | Snapdragon® X Elite | 2.619 ms | 148 - 148 MB | NPU
| pose_estimator | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 2.078 ms | 1 - 28 MB | NPU
| pose_estimator | ONNX | float | Snapdragon® 8 Gen 1 Mobile | 5.545 ms | 1 - 27 MB | NPU
| pose_estimator | ONNX | float | Qualcomm® QCS8550 (Proxy) | 2.72 ms | 0 - 82 MB | NPU
| pose_estimator | ONNX | float | Qualcomm® QCS8450 | 5.545 ms | 1 - 27 MB | NPU
| pose_estimator | ONNX | float | Snapdragon® 8 Elite Mobile | 1.64 ms | 0 - 24 MB | NPU
| pose_estimator | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 1.331 ms | 1 - 24 MB | NPU
| pose_estimator | ONNX | float | Qualcomm® QCS9075 | 4.548 ms | 1 - 46 MB | NPU
| pose_estimator | ONNX | float | Qualcomm® QCS8750 | 1.64 ms | 0 - 24 MB | NPU
| pose_estimator | ONNX | float | Qualcomm® QCS7181 | 2.619 ms | 148 - 148 MB | NPU
| pose_estimator | QNN_DLC | float | Snapdragon® X2 Elite | 1.479 ms | 1 - 1 MB | NPU
| pose_estimator | QNN_DLC | float | Snapdragon® X Elite | 2.867 ms | 1 - 1 MB | NPU
| pose_estimator | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 2.236 ms | 0 - 41 MB | NPU
| pose_estimator | QNN_DLC | float | Snapdragon® 8 Gen 1 Mobile | 6.514 ms | 0 - 40 MB | NPU
| pose_estimator | QNN_DLC | float | Qualcomm® QCS8275 | 17.821 ms | 1 - 25 MB | NPU
| pose_estimator | QNN_DLC | float | Qualcomm® QCS8550 (Proxy) | 2.776 ms | 1 - 3 MB | NPU
| pose_estimator | QNN_DLC | float | Qualcomm® QCS8450 | 6.514 ms | 0 - 40 MB | NPU
| pose_estimator | QNN_DLC | float | Snapdragon® 8 Elite Mobile | 1.736 ms | 1 - 27 MB | NPU
| pose_estimator | QNN_DLC | float | Qualcomm® SA8295P | 5.39 ms | 0 - 26 MB | NPU
| pose_estimator | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 1.424 ms | 1 - 29 MB | NPU
| pose_estimator | QNN_DLC | float | Qualcomm® SA7255P | 17.821 ms | 1 - 25 MB | NPU
| pose_estimator | QNN_DLC | float | Qualcomm® QCS9075 | 4.902 ms | 3 - 5 MB | NPU
| pose_estimator | QNN_DLC | float | Qualcomm® QCS8750 | 1.736 ms | 1 - 27 MB | NPU
| pose_estimator | QNN_DLC | float | Qualcomm® QCS7181 | 2.867 ms | 1 - 1 MB | NPU
| pose_estimator | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 2.249 ms | 0 - 44 MB | NPU
| pose_estimator | TFLITE | float | Snapdragon® 8 Gen 1 Mobile | 6.49 ms | 0 - 46 MB | NPU
| pose_estimator | TFLITE | float | Qualcomm® QCS8275 | 17.388 ms | 0 - 26 MB | NPU
| pose_estimator | TFLITE | float | Qualcomm® QCS8550 (Proxy) | 2.852 ms | 0 - 2 MB | NPU
| pose_estimator | TFLITE | float | Qualcomm® SA8775P | 41.489 ms | 0 - 17 MB | GPU
| pose_estimator | TFLITE | float | Qualcomm® SA8650P | 41.489 ms | 0 - 17 MB | GPU
| pose_estimator | TFLITE | float | Qualcomm® SA8255P | 41.489 ms | 0 - 17 MB | GPU
| pose_estimator | TFLITE | float | Qualcomm® QCS8450 | 6.49 ms | 0 - 46 MB | NPU
| pose_estimator | TFLITE | float | Snapdragon® 8 Elite Mobile | 1.704 ms | 0 - 30 MB | NPU
| pose_estimator | TFLITE | float | Qualcomm® SA8295P | 5.356 ms | 0 - 29 MB | NPU
| pose_estimator | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 1.378 ms | 0 - 29 MB | NPU
| pose_estimator | TFLITE | float | Qualcomm® SA7255P | 17.388 ms | 0 - 26 MB | NPU
| pose_estimator | TFLITE | float | Qualcomm® QCS9075 | 4.739 ms | 0 - 78 MB | NPU
| pose_estimator | TFLITE | float | Qualcomm® QCS8750 | 1.704 ms | 0 - 30 MB | NPU

## License
* The license for the original implementation of SixDRepNet can be found
  [here](https://github.com/thohemp/6DRepNet/blob/master/LICENSE).

## References
* [6D Rotation Representation for Unconstrained Head Pose Estimation](https://arxiv.org/abs/2109.10948)
* [Source Model Implementation](https://github.com/thohemp/6DRepNet)

## 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).