MediaPipe-Hand-Detection: Optimized for Mobile Deployment

Real-time hand detection optimized for mobile and edge

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

This model is an implementation of MediaPipe-Hand-Detection found here.

This repository provides scripts to run MediaPipe-Hand-Detection on Qualcomm® devices. More details on model performance across various devices, can be found here.

Model Details

Model Type: Object detection
Model Stats:
- Input resolution: 256x256
- Number of parameters (MediaPipeHandDetector): 1.76M
- Model size (MediaPipeHandDetector): 6.76 MB
- Number of parameters (MediaPipeHandLandmarkDetector): 2.01M
- Model size (MediaPipeHandLandmarkDetector): 7.71 MB

Model	Device	Chipset	Target Runtime	Inference Time (ms)	Peak Memory Range (MB)	Precision	Primary Compute Unit	Target Model
MediaPipeHandDetector	Samsung Galaxy S23	Snapdragon® 8 Gen 2	TFLITE	0.722 ms	0 - 22 MB	FP16	NPU	MediaPipe-Hand-Detection.tflite
MediaPipeHandDetector	Samsung Galaxy S23	Snapdragon® 8 Gen 2	ONNX	1.182 ms	0 - 6 MB	FP16	NPU	MediaPipe-Hand-Detection.onnx
MediaPipeHandDetector	Samsung Galaxy S24	Snapdragon® 8 Gen 3	TFLITE	0.518 ms	0 - 19 MB	FP16	NPU	MediaPipe-Hand-Detection.tflite
MediaPipeHandDetector	Samsung Galaxy S24	Snapdragon® 8 Gen 3	ONNX	0.859 ms	0 - 70 MB	FP16	NPU	MediaPipe-Hand-Detection.onnx
MediaPipeHandDetector	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	TFLITE	0.523 ms	0 - 16 MB	FP16	NPU	MediaPipe-Hand-Detection.tflite
MediaPipeHandDetector	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	ONNX	0.728 ms	0 - 36 MB	FP16	NPU	MediaPipe-Hand-Detection.onnx
MediaPipeHandDetector	QCS8550 (Proxy)	QCS8550 Proxy	TFLITE	0.713 ms	0 - 75 MB	FP16	NPU	MediaPipe-Hand-Detection.tflite
MediaPipeHandDetector	SA7255P ADP	SA7255P	TFLITE	24.782 ms	0 - 17 MB	FP16	NPU	MediaPipe-Hand-Detection.tflite
MediaPipeHandDetector	SA8255 (Proxy)	SA8255P Proxy	TFLITE	0.716 ms	0 - 22 MB	FP16	NPU	MediaPipe-Hand-Detection.tflite
MediaPipeHandDetector	SA8295P ADP	SA8295P	TFLITE	1.754 ms	0 - 11 MB	FP16	NPU	MediaPipe-Hand-Detection.tflite
MediaPipeHandDetector	SA8650 (Proxy)	SA8650P Proxy	TFLITE	0.723 ms	0 - 22 MB	FP16	NPU	MediaPipe-Hand-Detection.tflite
MediaPipeHandDetector	SA8775P ADP	SA8775P	TFLITE	1.526 ms	0 - 16 MB	FP16	NPU	MediaPipe-Hand-Detection.tflite
MediaPipeHandDetector	QCS8450 (Proxy)	QCS8450 Proxy	TFLITE	1.292 ms	0 - 19 MB	FP16	NPU	MediaPipe-Hand-Detection.tflite
MediaPipeHandDetector	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	1.19 ms	4 - 4 MB	FP16	NPU	MediaPipe-Hand-Detection.onnx
MediaPipeHandLandmarkDetector	Samsung Galaxy S23	Snapdragon® 8 Gen 2	TFLITE	1.019 ms	0 - 36 MB	FP16	NPU	MediaPipe-Hand-Detection.tflite
MediaPipeHandLandmarkDetector	Samsung Galaxy S23	Snapdragon® 8 Gen 2	ONNX	1.544 ms	0 - 8 MB	FP16	NPU	MediaPipe-Hand-Detection.onnx
MediaPipeHandLandmarkDetector	Samsung Galaxy S24	Snapdragon® 8 Gen 3	TFLITE	0.731 ms	0 - 18 MB	FP16	NPU	MediaPipe-Hand-Detection.tflite
MediaPipeHandLandmarkDetector	Samsung Galaxy S24	Snapdragon® 8 Gen 3	ONNX	1.135 ms	0 - 69 MB	FP16	NPU	MediaPipe-Hand-Detection.onnx
MediaPipeHandLandmarkDetector	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	TFLITE	0.698 ms	0 - 17 MB	FP16	NPU	MediaPipe-Hand-Detection.tflite
MediaPipeHandLandmarkDetector	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	ONNX	1.066 ms	0 - 40 MB	FP16	NPU	MediaPipe-Hand-Detection.onnx
MediaPipeHandLandmarkDetector	QCS8550 (Proxy)	QCS8550 Proxy	TFLITE	1.001 ms	0 - 6 MB	FP16	NPU	MediaPipe-Hand-Detection.tflite
MediaPipeHandLandmarkDetector	SA7255P ADP	SA7255P	TFLITE	35.495 ms	0 - 16 MB	FP16	NPU	MediaPipe-Hand-Detection.tflite
MediaPipeHandLandmarkDetector	SA8255 (Proxy)	SA8255P Proxy	TFLITE	1.018 ms	0 - 47 MB	FP16	NPU	MediaPipe-Hand-Detection.tflite
MediaPipeHandLandmarkDetector	SA8295P ADP	SA8295P	TFLITE	2.271 ms	0 - 12 MB	FP16	NPU	MediaPipe-Hand-Detection.tflite
MediaPipeHandLandmarkDetector	SA8650 (Proxy)	SA8650P Proxy	TFLITE	1.023 ms	0 - 229 MB	FP16	NPU	MediaPipe-Hand-Detection.tflite
MediaPipeHandLandmarkDetector	SA8775P ADP	SA8775P	TFLITE	2.213 ms	0 - 17 MB	FP16	NPU	MediaPipe-Hand-Detection.tflite
MediaPipeHandLandmarkDetector	QCS8450 (Proxy)	QCS8450 Proxy	TFLITE	1.795 ms	0 - 16 MB	FP16	NPU	MediaPipe-Hand-Detection.tflite
MediaPipeHandLandmarkDetector	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	1.649 ms	6 - 6 MB	FP16	NPU	MediaPipe-Hand-Detection.onnx

Installation

This model can be installed as a Python package via pip.

pip install qai-hub-models

Configure Qualcomm® AI Hub to run this model on a cloud-hosted device

Sign-in to Qualcomm® AI Hub with your Qualcomm® ID. Once signed in navigate to Account -> Settings -> API Token.

With this API token, you can configure your client to run models on the cloud hosted devices.

qai-hub configure --api_token API_TOKEN

Navigate to docs for more information.

Demo off target

The package contains a simple end-to-end demo that downloads pre-trained weights and runs this model on a sample input.

python -m qai_hub_models.models.mediapipe_hand.demo

The above demo runs a reference implementation of pre-processing, model inference, and post processing.

NOTE: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above).

%run -m qai_hub_models.models.mediapipe_hand.demo

Run model on a cloud-hosted device

In addition to the demo, you can also run the model on a cloud-hosted Qualcomm® device. This script does the following:

Performance check on-device on a cloud-hosted device
Downloads compiled assets that can be deployed on-device for Android.
Accuracy check between PyTorch and on-device outputs.

python -m qai_hub_models.models.mediapipe_hand.export

Profiling Results
------------------------------------------------------------
MediaPipeHandDetector
Device                          : Samsung Galaxy S23 (13)
Runtime                         : TFLITE                 
Estimated inference time (ms)   : 0.7                    
Estimated peak memory usage (MB): [0, 22]                
Total # Ops                     : 149                    
Compute Unit(s)                 : NPU (149 ops)          

------------------------------------------------------------
MediaPipeHandLandmarkDetector
Device                          : Samsung Galaxy S23 (13)
Runtime                         : TFLITE                 
Estimated inference time (ms)   : 1.0                    
Estimated peak memory usage (MB): [0, 36]                
Total # Ops                     : 158                    
Compute Unit(s)                 : NPU (158 ops)

How does this work?

This export script leverages Qualcomm® AI Hub to optimize, validate, and deploy this model on-device. Lets go through each step below in detail:

Step 1: Compile model for on-device deployment

To compile a PyTorch model for on-device deployment, we first trace the model in memory using the jit.trace and then call the submit_compile_job API.

import torch

import qai_hub as hub
from qai_hub_models.models.mediapipe_hand import Model

# Load the model
model = Model.from_pretrained()
hand_detector_model = model.hand_detector
hand_landmark_detector_model = model.hand_landmark_detector

# Device
device = hub.Device("Samsung Galaxy S23")

# Trace model
hand_detector_input_shape = hand_detector_model.get_input_spec()
hand_detector_sample_inputs = hand_detector_model.sample_inputs()

traced_hand_detector_model = torch.jit.trace(hand_detector_model, [torch.tensor(data[0]) for _, data in hand_detector_sample_inputs.items()])

# Compile model on a specific device
hand_detector_compile_job = hub.submit_compile_job(
    model=traced_hand_detector_model ,
    device=device,
    input_specs=hand_detector_model.get_input_spec(),
)

# Get target model to run on-device
hand_detector_target_model = hand_detector_compile_job.get_target_model()
# Trace model
hand_landmark_detector_input_shape = hand_landmark_detector_model.get_input_spec()
hand_landmark_detector_sample_inputs = hand_landmark_detector_model.sample_inputs()

traced_hand_landmark_detector_model = torch.jit.trace(hand_landmark_detector_model, [torch.tensor(data[0]) for _, data in hand_landmark_detector_sample_inputs.items()])

# Compile model on a specific device
hand_landmark_detector_compile_job = hub.submit_compile_job(
    model=traced_hand_landmark_detector_model ,
    device=device,
    input_specs=hand_landmark_detector_model.get_input_spec(),
)

# Get target model to run on-device
hand_landmark_detector_target_model = hand_landmark_detector_compile_job.get_target_model()

Step 2: Performance profiling on cloud-hosted device

After compiling models from step 1. Models can be profiled model on-device using the target_model. Note that this scripts runs the model on a device automatically provisioned in the cloud. Once the job is submitted, you can navigate to a provided job URL to view a variety of on-device performance metrics.

hand_detector_profile_job = hub.submit_profile_job(
    model=hand_detector_target_model,
    device=device,
)
hand_landmark_detector_profile_job = hub.submit_profile_job(
    model=hand_landmark_detector_target_model,
    device=device,
)

Step 3: Verify on-device accuracy

To verify the accuracy of the model on-device, you can run on-device inference on sample input data on the same cloud hosted device.

hand_detector_input_data = hand_detector_model.sample_inputs()
hand_detector_inference_job = hub.submit_inference_job(
    model=hand_detector_target_model,
    device=device,
    inputs=hand_detector_input_data,
)
hand_detector_inference_job.download_output_data()
hand_landmark_detector_input_data = hand_landmark_detector_model.sample_inputs()
hand_landmark_detector_inference_job = hub.submit_inference_job(
    model=hand_landmark_detector_target_model,
    device=device,
    inputs=hand_landmark_detector_input_data,
)
hand_landmark_detector_inference_job.download_output_data()

With the output of the model, you can compute like PSNR, relative errors or spot check the output with expected output.

Note: This on-device profiling and inference requires access to Qualcomm® AI Hub. Sign up for access.

Deploying compiled model to Android

The models can be deployed using multiple runtimes:

TensorFlow Lite (.tflite export): This tutorial provides a guide to deploy the .tflite model in an Android application.
QNN (.so export ): This sample app provides instructions on how to use the .so shared library in an Android application.

View on Qualcomm® AI Hub

Get more details on MediaPipe-Hand-Detection's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

The license for the original implementation of MediaPipe-Hand-Detection can be found here.
The license for the compiled assets for on-device deployment can be found here

References

Community

Join our AI Hub Slack community to collaborate, post questions and learn more about on-device AI.
For questions or feedback please reach out to us.