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

---

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

# MediaPipe-Hand-Detection: Optimized for Qualcomm Devices

The MediaPipe Hand Landmark Detector is a machine learning pipeline that predicts bounding boxes and pose skeletons of hands in an image.

This is based on the implementation of MediaPipe-Hand-Detection found [here](https://github.com/zmurez/MediaPipePyTorch/).
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/main/src/qai_hub_models/models/mediapipe_hand) 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 2.42, ONNX Runtime 1.24.3 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/mediapipe_hand/releases/v0.51.0/mediapipe_hand-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/mediapipe_hand/releases/v0.51.0/mediapipe_hand-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/mediapipe_hand/releases/v0.51.0/mediapipe_hand-tflite-float.zip)

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


### Option 2: Export with Custom Configurations

Use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/main/src/qai_hub_models/models/mediapipe_hand) 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 [MediaPipe-Hand-Detection on GitHub](https://github.com/qualcomm/ai-hub-models/blob/main/src/qai_hub_models/models/mediapipe_hand) for usage instructions.

## Model Details

**Model Type:** Model_use_case.object_detection

**Model Stats:**
- Input resolution: 256x256
- Number of parameters (hand_detector): 1.76M
- Model size (hand_detector) (float): 6.75 MB
- Number of parameters (hand_landmark_detector): 2.01M
- Model size (hand_landmark_detector) (float): 7.70 MB

## Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit
|---|---|---|---|---|---|---
| hand_detector | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.418 ms | 0 - 47 MB | NPU
| hand_detector | ONNX | float | Snapdragon® X2 Elite | 0.493 ms | 0 - 0 MB | NPU
| hand_detector | ONNX | float | Snapdragon® X Elite | 1.026 ms | 3 - 3 MB | NPU
| hand_detector | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 0.594 ms | 0 - 73 MB | NPU
| hand_detector | ONNX | float | Qualcomm® QCS8550 (Proxy) | 0.851 ms | 0 - 110 MB | NPU
| hand_detector | ONNX | float | Qualcomm® QCS9075 | 1.271 ms | 1 - 3 MB | NPU
| hand_detector | ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 0.497 ms | 0 - 46 MB | NPU
| hand_detector | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.377 ms | 1 - 42 MB | NPU
| hand_detector | QNN_DLC | float | Snapdragon® X2 Elite | 0.606 ms | 1 - 1 MB | NPU
| hand_detector | QNN_DLC | float | Snapdragon® X Elite | 0.889 ms | 1 - 1 MB | NPU
| hand_detector | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 0.533 ms | 0 - 57 MB | NPU
| hand_detector | QNN_DLC | float | Qualcomm® QCS8275 (Proxy) | 3.797 ms | 1 - 39 MB | NPU
| hand_detector | QNN_DLC | float | Qualcomm® QCS8550 (Proxy) | 0.731 ms | 1 - 2 MB | NPU
| hand_detector | QNN_DLC | float | Qualcomm® SA8775P | 1.272 ms | 0 - 41 MB | NPU
| hand_detector | QNN_DLC | float | Qualcomm® QCS9075 | 1.116 ms | 1 - 3 MB | NPU
| hand_detector | QNN_DLC | float | Qualcomm® QCS8450 (Proxy) | 1.376 ms | 0 - 53 MB | NPU
| hand_detector | QNN_DLC | float | Qualcomm® SA7255P | 3.797 ms | 1 - 39 MB | NPU
| hand_detector | QNN_DLC | float | Qualcomm® SA8295P | 1.693 ms | 0 - 31 MB | NPU
| hand_detector | QNN_DLC | float | Snapdragon® 8 Elite For Galaxy Mobile | 0.441 ms | 0 - 43 MB | NPU
| hand_detector | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.38 ms | 0 - 41 MB | NPU
| hand_detector | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 0.536 ms | 0 - 55 MB | NPU
| hand_detector | TFLITE | float | Qualcomm® QCS8275 (Proxy) | 3.802 ms | 0 - 37 MB | NPU
| hand_detector | TFLITE | float | Qualcomm® QCS8550 (Proxy) | 0.734 ms | 0 - 4 MB | NPU
| hand_detector | TFLITE | float | Qualcomm® SA8775P | 1.291 ms | 0 - 40 MB | NPU
| hand_detector | TFLITE | float | Qualcomm® QCS9075 | 1.143 ms | 0 - 7 MB | NPU
| hand_detector | TFLITE | float | Qualcomm® QCS8450 (Proxy) | 1.376 ms | 0 - 50 MB | NPU
| hand_detector | TFLITE | float | Qualcomm® SA7255P | 3.802 ms | 0 - 37 MB | NPU
| hand_detector | TFLITE | float | Qualcomm® SA8295P | 1.706 ms | 0 - 30 MB | NPU
| hand_detector | TFLITE | float | Snapdragon® 8 Elite For Galaxy Mobile | 0.445 ms | 0 - 41 MB | NPU
| hand_landmark_detector | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.566 ms | 0 - 43 MB | NPU
| hand_landmark_detector | ONNX | float | Snapdragon® X2 Elite | 0.722 ms | 6 - 6 MB | NPU
| hand_landmark_detector | ONNX | float | Snapdragon® X Elite | 1.377 ms | 6 - 6 MB | NPU
| hand_landmark_detector | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 0.838 ms | 0 - 59 MB | NPU
| hand_landmark_detector | ONNX | float | Qualcomm® QCS8550 (Proxy) | 1.103 ms | 0 - 8 MB | NPU
| hand_landmark_detector | ONNX | float | Qualcomm® QCS9075 | 1.824 ms | 1 - 4 MB | NPU
| hand_landmark_detector | ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 0.658 ms | 0 - 43 MB | NPU
| hand_landmark_detector | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.509 ms | 1 - 39 MB | NPU
| hand_landmark_detector | QNN_DLC | float | Snapdragon® X2 Elite | 0.855 ms | 1 - 1 MB | NPU
| hand_landmark_detector | QNN_DLC | float | Snapdragon® X Elite | 1.285 ms | 1 - 1 MB | NPU
| hand_landmark_detector | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 0.748 ms | 0 - 56 MB | NPU
| hand_landmark_detector | QNN_DLC | float | Qualcomm® QCS8275 (Proxy) | 5.325 ms | 1 - 35 MB | NPU
| hand_landmark_detector | QNN_DLC | float | Qualcomm® QCS8550 (Proxy) | 0.988 ms | 1 - 2 MB | NPU
| hand_landmark_detector | QNN_DLC | float | Qualcomm® SA8775P | 1.861 ms | 0 - 37 MB | NPU
| hand_landmark_detector | QNN_DLC | float | Qualcomm® QCS9075 | 1.69 ms | 1 - 3 MB | NPU
| hand_landmark_detector | QNN_DLC | float | Qualcomm® QCS8450 (Proxy) | 1.961 ms | 0 - 52 MB | NPU
| hand_landmark_detector | QNN_DLC | float | Qualcomm® SA7255P | 5.325 ms | 1 - 35 MB | NPU
| hand_landmark_detector | QNN_DLC | float | Qualcomm® SA8295P | 2.238 ms | 1 - 31 MB | NPU
| hand_landmark_detector | QNN_DLC | float | Snapdragon® 8 Elite For Galaxy Mobile | 0.602 ms | 1 - 39 MB | NPU
| hand_landmark_detector | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.504 ms | 0 - 43 MB | NPU
| hand_landmark_detector | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 0.76 ms | 0 - 59 MB | NPU
| hand_landmark_detector | TFLITE | float | Qualcomm® QCS8275 (Proxy) | 5.357 ms | 0 - 39 MB | NPU
| hand_landmark_detector | TFLITE | float | Qualcomm® QCS8550 (Proxy) | 1.076 ms | 0 - 2 MB | NPU
| hand_landmark_detector | TFLITE | float | Qualcomm® SA8775P | 1.85 ms | 0 - 43 MB | NPU
| hand_landmark_detector | TFLITE | float | Qualcomm® QCS9075 | 1.681 ms | 0 - 9 MB | NPU
| hand_landmark_detector | TFLITE | float | Qualcomm® QCS8450 (Proxy) | 1.949 ms | 0 - 56 MB | NPU
| hand_landmark_detector | TFLITE | float | Qualcomm® SA7255P | 5.357 ms | 0 - 39 MB | NPU
| hand_landmark_detector | TFLITE | float | Qualcomm® SA8295P | 2.232 ms | 0 - 34 MB | NPU
| hand_landmark_detector | TFLITE | float | Snapdragon® 8 Elite For Galaxy Mobile | 0.606 ms | 0 - 37 MB | NPU

## License
* The license for the original implementation of MediaPipe-Hand-Detection can be found
  [here](https://github.com/zmurez/MediaPipePyTorch/blob/master/LICENSE).

## References
* [MediaPipe Hands: On-device Real-time Hand Tracking](https://arxiv.org/abs/2006.10214)
* [Source Model Implementation](https://github.com/zmurez/MediaPipePyTorch/)

## 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).