MediaPipe-Pose-Estimation: Optimized for Mobile Deployment

Detect and track human body poses in real-time images and video streams

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

This model is an implementation of MediaPipe-Pose-Estimation found here.

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

Model Details

• Model Type: Pose estimation
• Model Stats:
  • Input resolution: 256x256
  • Number of parameters (MediaPipePoseDetector): 815K
  • Model size (MediaPipePoseDetector): 3.14 MB
  • Number of parameters (MediaPipePoseLandmarkDetector): 3.37M
  • Model size (MediaPipePoseLandmarkDetector): 12.9 MB

Model	Device	Chipset	Target Runtime	Inference Time (ms)	Peak Memory Range (MB)	Precision	Primary Compute Unit	Target Model
MediaPipePoseDetector	Samsung Galaxy S23	Snapdragon® 8 Gen 2	TFLITE	0.777 ms	0 - 22 MB	FP16	NPU	MediaPipe-Pose-Estimation.tflite
MediaPipePoseDetector	Samsung Galaxy S23	Snapdragon® 8 Gen 2	ONNX	0.997 ms	0 - 4 MB	FP16	NPU	MediaPipe-Pose-Estimation.onnx
MediaPipePoseDetector	Samsung Galaxy S24	Snapdragon® 8 Gen 3	TFLITE	0.56 ms	0 - 17 MB	FP16	NPU	MediaPipe-Pose-Estimation.tflite
MediaPipePoseDetector	Samsung Galaxy S24	Snapdragon® 8 Gen 3	ONNX	0.744 ms	0 - 50 MB	FP16	NPU	MediaPipe-Pose-Estimation.onnx
MediaPipePoseDetector	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	TFLITE	0.57 ms	0 - 12 MB	FP16	NPU	MediaPipe-Pose-Estimation.tflite
MediaPipePoseDetector	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	ONNX	0.756 ms	0 - 26 MB	FP16	NPU	MediaPipe-Pose-Estimation.onnx
MediaPipePoseDetector	QCS8550 (Proxy)	QCS8550 Proxy	TFLITE	0.77 ms	0 - 58 MB	FP16	NPU	MediaPipe-Pose-Estimation.tflite
MediaPipePoseDetector	SA7255P ADP	SA7255P	TFLITE	37.961 ms	0 - 13 MB	FP16	NPU	MediaPipe-Pose-Estimation.tflite
MediaPipePoseDetector	SA8255 (Proxy)	SA8255P Proxy	TFLITE	0.772 ms	0 - 7 MB	FP16	NPU	MediaPipe-Pose-Estimation.tflite
MediaPipePoseDetector	SA8295P ADP	SA8295P	TFLITE	2.335 ms	0 - 10 MB	FP16	NPU	MediaPipe-Pose-Estimation.tflite
MediaPipePoseDetector	SA8650 (Proxy)	SA8650P Proxy	TFLITE	0.772 ms	0 - 91 MB	FP16	NPU	MediaPipe-Pose-Estimation.tflite
MediaPipePoseDetector	SA8775P ADP	SA8775P	TFLITE	1.784 ms	0 - 11 MB	FP16	NPU	MediaPipe-Pose-Estimation.tflite
MediaPipePoseDetector	QCS8450 (Proxy)	QCS8450 Proxy	TFLITE	1.9 ms	0 - 14 MB	FP16	NPU	MediaPipe-Pose-Estimation.tflite
MediaPipePoseDetector	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	1.073 ms	4 - 4 MB	FP16	NPU	MediaPipe-Pose-Estimation.onnx
MediaPipePoseLandmarkDetector	Samsung Galaxy S23	Snapdragon® 8 Gen 2	TFLITE	0.821 ms	0 - 44 MB	FP16	NPU	MediaPipe-Pose-Estimation.tflite
MediaPipePoseLandmarkDetector	Samsung Galaxy S23	Snapdragon® 8 Gen 2	ONNX	1.309 ms	0 - 44 MB	FP16	NPU	MediaPipe-Pose-Estimation.onnx
MediaPipePoseLandmarkDetector	Samsung Galaxy S24	Snapdragon® 8 Gen 3	TFLITE	0.603 ms	0 - 22 MB	FP16	NPU	MediaPipe-Pose-Estimation.tflite
MediaPipePoseLandmarkDetector	Samsung Galaxy S24	Snapdragon® 8 Gen 3	ONNX	0.968 ms	0 - 98 MB	FP16	NPU	MediaPipe-Pose-Estimation.onnx
MediaPipePoseLandmarkDetector	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	TFLITE	0.553 ms	0 - 19 MB	FP16	NPU	MediaPipe-Pose-Estimation.tflite
MediaPipePoseLandmarkDetector	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	ONNX	0.912 ms	0 - 45 MB	FP16	NPU	MediaPipe-Pose-Estimation.onnx
MediaPipePoseLandmarkDetector	QCS8550 (Proxy)	QCS8550 Proxy	TFLITE	0.801 ms	0 - 44 MB	FP16	NPU	MediaPipe-Pose-Estimation.tflite
MediaPipePoseLandmarkDetector	SA7255P ADP	SA7255P	TFLITE	17.243 ms	0 - 19 MB	FP16	NPU	MediaPipe-Pose-Estimation.tflite
MediaPipePoseLandmarkDetector	SA8255 (Proxy)	SA8255P Proxy	TFLITE	0.803 ms	0 - 11 MB	FP16	NPU	MediaPipe-Pose-Estimation.tflite
MediaPipePoseLandmarkDetector	SA8295P ADP	SA8295P	TFLITE	1.417 ms	0 - 14 MB	FP16	NPU	MediaPipe-Pose-Estimation.tflite
MediaPipePoseLandmarkDetector	SA8650 (Proxy)	SA8650P Proxy	TFLITE	0.832 ms	0 - 234 MB	FP16	NPU	MediaPipe-Pose-Estimation.tflite
MediaPipePoseLandmarkDetector	SA8775P ADP	SA8775P	TFLITE	1.629 ms	0 - 20 MB	FP16	NPU	MediaPipe-Pose-Estimation.tflite
MediaPipePoseLandmarkDetector	QCS8450 (Proxy)	QCS8450 Proxy	TFLITE	1.004 ms	0 - 17 MB	FP16	NPU	MediaPipe-Pose-Estimation.tflite
MediaPipePoseLandmarkDetector	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	1.389 ms	8 - 8 MB	FP16	NPU	MediaPipe-Pose-Estimation.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_pose.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_pose.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_pose.export

Profiling Results
------------------------------------------------------------
MediaPipePoseDetector
Device                          : Samsung Galaxy S23 (13)
Runtime                         : TFLITE                 
Estimated inference time (ms)   : 0.8                    
Estimated peak memory usage (MB): [0, 22]                
Total # Ops                     : 106                    
Compute Unit(s)                 : NPU (106 ops)          

------------------------------------------------------------
MediaPipePoseLandmarkDetector
Device                          : Samsung Galaxy S23 (13)
Runtime                         : TFLITE                 
Estimated inference time (ms)   : 0.8                    
Estimated peak memory usage (MB): [0, 44]                
Total # Ops                     : 219                    
Compute Unit(s)                 : NPU (219 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_pose import Model

# Load the model
model = Model.from_pretrained()
pose_detector_model = model.pose_detector
pose_landmark_detector_model = model.pose_landmark_detector

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

# Trace model
pose_detector_input_shape = pose_detector_model.get_input_spec()
pose_detector_sample_inputs = pose_detector_model.sample_inputs()

traced_pose_detector_model = torch.jit.trace(pose_detector_model, [torch.tensor(data[0]) for _, data in pose_detector_sample_inputs.items()])

# Compile model on a specific device
pose_detector_compile_job = hub.submit_compile_job(
    model=traced_pose_detector_model ,
    device=device,
    input_specs=pose_detector_model.get_input_spec(),
)

# Get target model to run on-device
pose_detector_target_model = pose_detector_compile_job.get_target_model()
# Trace model
pose_landmark_detector_input_shape = pose_landmark_detector_model.get_input_spec()
pose_landmark_detector_sample_inputs = pose_landmark_detector_model.sample_inputs()

traced_pose_landmark_detector_model = torch.jit.trace(pose_landmark_detector_model, [torch.tensor(data[0]) for _, data in pose_landmark_detector_sample_inputs.items()])

# Compile model on a specific device
pose_landmark_detector_compile_job = hub.submit_compile_job(
    model=traced_pose_landmark_detector_model ,
    device=device,
    input_specs=pose_landmark_detector_model.get_input_spec(),
)

# Get target model to run on-device
pose_landmark_detector_target_model = pose_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.

pose_detector_profile_job = hub.submit_profile_job(
    model=pose_detector_target_model,
    device=device,
)
pose_landmark_detector_profile_job = hub.submit_profile_job(
    model=pose_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.

pose_detector_input_data = pose_detector_model.sample_inputs()
pose_detector_inference_job = hub.submit_inference_job(
    model=pose_detector_target_model,
    device=device,
    inputs=pose_detector_input_data,
)
pose_detector_inference_job.download_output_data()
pose_landmark_detector_input_data = pose_landmark_detector_model.sample_inputs()
pose_landmark_detector_inference_job = hub.submit_inference_job(
    model=pose_landmark_detector_target_model,
    device=device,
    inputs=pose_landmark_detector_input_data,
)
pose_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-Pose-Estimation's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

• The license for the original implementation of MediaPipe-Pose-Estimation can be found here.
• The license for the compiled assets for on-device deployment can be found here

References

• BlazePose: On-device Real-time Body Pose tracking
• Source Model Implementation

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.