Person-Foot-Detection: Optimized for Mobile Deployment

Multi-task Human detector

Real-time multiple person detection with accurate feet localization optimized for mobile and edge.

This model is an implementation of Person-Foot-Detection found here.

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

Model Details

• Model Type: Model_use_case.object_detection
• Model Stats:
  • Inference latency: RealTime
  • Input resolution: 640x480
  • Number of output classes: 2
  • Number of parameters: 2.53M
  • Model size (float): 9.69 MB
  • Model size (w8a8): 2.62 MB
  • Model size (w8a16): 2.90 MB

Model	Precision	Device	Chipset	Target Runtime	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit	Target Model
Person-Foot-Detection	float	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	TFLITE	23.186 ms	5 - 28 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	float	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	QNN_DLC	22.212 ms	4 - 27 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	float	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	TFLITE	7.028 ms	5 - 37 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	float	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	QNN_DLC	9.165 ms	4 - 41 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	TFLITE	4.847 ms	5 - 14 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	QNN_DLC	4.46 ms	4 - 10 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	float	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	TFLITE	7.72 ms	5 - 29 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	float	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	QNN_DLC	7.267 ms	2 - 27 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	float	SA7255P ADP	Qualcomm® SA7255P	TFLITE	23.186 ms	5 - 28 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	float	SA7255P ADP	Qualcomm® SA7255P	QNN_DLC	22.212 ms	4 - 27 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	float	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	TFLITE	4.92 ms	5 - 15 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	float	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	QNN_DLC	4.453 ms	4 - 10 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	float	SA8295P ADP	Qualcomm® SA8295P	TFLITE	8.343 ms	5 - 30 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	float	SA8295P ADP	Qualcomm® SA8295P	QNN_DLC	7.935 ms	0 - 32 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	float	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	TFLITE	4.894 ms	2 - 12 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	float	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	QNN_DLC	4.456 ms	4 - 11 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	float	SA8775P ADP	Qualcomm® SA8775P	TFLITE	7.72 ms	5 - 29 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	float	SA8775P ADP	Qualcomm® SA8775P	QNN_DLC	7.267 ms	2 - 27 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	float	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	TFLITE	4.897 ms	4 - 14 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	float	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	QNN_DLC	4.546 ms	4 - 10 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	float	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	ONNX	5.454 ms	16 - 36 MB	NPU	Person-Foot-Detection.onnx
Person-Foot-Detection	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	TFLITE	3.379 ms	0 - 33 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	QNN_DLC	3.14 ms	4 - 39 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	ONNX	3.812 ms	2 - 43 MB	NPU	Person-Foot-Detection.onnx
Person-Foot-Detection	float	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	TFLITE	3.487 ms	0 - 28 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	float	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	QNN_DLC	3.177 ms	4 - 33 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	float	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	ONNX	3.767 ms	16 - 47 MB	NPU	Person-Foot-Detection.onnx
Person-Foot-Detection	float	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN_DLC	5.347 ms	10 - 10 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	float	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	5.405 ms	20 - 20 MB	NPU	Person-Foot-Detection.onnx
Person-Foot-Detection	w8a16	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	QNN_DLC	8.463 ms	2 - 29 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a16	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	QNN_DLC	5.413 ms	2 - 42 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a16	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	QNN_DLC	3.537 ms	2 - 10 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a16	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	QNN_DLC	4.292 ms	2 - 30 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a16	RB3 Gen 2 (Proxy)	Qualcomm® QCS6490 (Proxy)	QNN_DLC	17.902 ms	2 - 30 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a16	SA7255P ADP	Qualcomm® SA7255P	QNN_DLC	8.463 ms	2 - 29 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a16	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	QNN_DLC	3.546 ms	2 - 10 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a16	SA8295P ADP	Qualcomm® SA8295P	QNN_DLC	5.275 ms	2 - 30 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a16	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	QNN_DLC	3.565 ms	2 - 10 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a16	SA8775P ADP	Qualcomm® SA8775P	QNN_DLC	4.292 ms	2 - 30 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a16	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	QNN_DLC	3.551 ms	2 - 9 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a16	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	ONNX	4.304 ms	8 - 19 MB	NPU	Person-Foot-Detection.onnx
Person-Foot-Detection	w8a16	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	QNN_DLC	2.257 ms	2 - 38 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a16	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	ONNX	2.902 ms	0 - 45 MB	NPU	Person-Foot-Detection.onnx
Person-Foot-Detection	w8a16	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	QNN_DLC	1.923 ms	2 - 35 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a16	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	ONNX	2.813 ms	9 - 45 MB	NPU	Person-Foot-Detection.onnx
Person-Foot-Detection	w8a16	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN_DLC	4.703 ms	1 - 1 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a16	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	4.712 ms	9 - 9 MB	NPU	Person-Foot-Detection.onnx
Person-Foot-Detection	w8a8	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	TFLITE	3.903 ms	1 - 23 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	w8a8	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	QNN_DLC	3.472 ms	1 - 26 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a8	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	TFLITE	1.53 ms	0 - 35 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	w8a8	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	QNN_DLC	1.9 ms	1 - 39 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a8	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	TFLITE	1.31 ms	0 - 12 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	w8a8	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	QNN_DLC	1.128 ms	1 - 6 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a8	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	TFLITE	1.73 ms	0 - 24 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	w8a8	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	QNN_DLC	1.544 ms	1 - 28 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a8	RB3 Gen 2 (Proxy)	Qualcomm® QCS6490 (Proxy)	TFLITE	5.546 ms	1 - 30 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	w8a8	RB3 Gen 2 (Proxy)	Qualcomm® QCS6490 (Proxy)	QNN_DLC	6.861 ms	1 - 30 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a8	RB5 (Proxy)	Qualcomm® QCS8250 (Proxy)	TFLITE	25.879 ms	1 - 4 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	w8a8	SA7255P ADP	Qualcomm® SA7255P	TFLITE	3.903 ms	1 - 23 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	w8a8	SA7255P ADP	Qualcomm® SA7255P	QNN_DLC	3.472 ms	1 - 26 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a8	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	TFLITE	1.27 ms	0 - 12 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	w8a8	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	QNN_DLC	1.14 ms	1 - 12 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a8	SA8295P ADP	Qualcomm® SA8295P	TFLITE	2.373 ms	0 - 24 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	w8a8	SA8295P ADP	Qualcomm® SA8295P	QNN_DLC	2.159 ms	1 - 32 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a8	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	TFLITE	1.304 ms	0 - 12 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	w8a8	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	QNN_DLC	1.137 ms	1 - 12 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a8	SA8775P ADP	Qualcomm® SA8775P	TFLITE	1.73 ms	0 - 24 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	w8a8	SA8775P ADP	Qualcomm® SA8775P	QNN_DLC	1.544 ms	1 - 28 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a8	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	TFLITE	1.276 ms	0 - 12 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	w8a8	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	QNN_DLC	1.129 ms	1 - 12 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a8	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	ONNX	1.946 ms	0 - 13 MB	NPU	Person-Foot-Detection.onnx
Person-Foot-Detection	w8a8	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	TFLITE	0.895 ms	0 - 42 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	w8a8	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	QNN_DLC	0.758 ms	1 - 38 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a8	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	ONNX	1.267 ms	0 - 46 MB	NPU	Person-Foot-Detection.onnx
Person-Foot-Detection	w8a8	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	TFLITE	0.879 ms	0 - 26 MB	NPU	Person-Foot-Detection.tflite
Person-Foot-Detection	w8a8	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	QNN_DLC	0.743 ms	0 - 27 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a8	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	ONNX	1.258 ms	0 - 40 MB	NPU	Person-Foot-Detection.onnx
Person-Foot-Detection	w8a8	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN_DLC	1.498 ms	1 - 1 MB	NPU	Person-Foot-Detection.dlc
Person-Foot-Detection	w8a8	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	2.103 ms	6 - 6 MB	NPU	Person-Foot-Detection.onnx

Installation

Install the package via pip:

pip install "qai-hub-models[foot-track-net]"

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.foot_track_net.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.foot_track_net.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.foot_track_net.export

Profiling Results
------------------------------------------------------------
Person-Foot-Detection
Device                          : cs_8275 (ANDROID 14)                 
Runtime                         : TFLITE                               
Estimated inference time (ms)   : 23.2                                 
Estimated peak memory usage (MB): [5, 28]                              
Total # Ops                     : 135                                  
Compute Unit(s)                 : npu (135 ops) gpu (0 ops) cpu (0 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.foot_track_net import Model

# Load the model
torch_model = Model.from_pretrained()

# Device
device = hub.Device("Samsung Galaxy S24")

# Trace model
input_shape = torch_model.get_input_spec()
sample_inputs = torch_model.sample_inputs()

pt_model = torch.jit.trace(torch_model, [torch.tensor(data[0]) for _, data in sample_inputs.items()])

# Compile model on a specific device
compile_job = hub.submit_compile_job(
    model=pt_model,
    device=device,
    input_specs=torch_model.get_input_spec(),
)

# Get target model to run on-device
target_model = 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.

profile_job = hub.submit_profile_job(
    model=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.

input_data = torch_model.sample_inputs()
inference_job = hub.submit_inference_job(
    model=target_model,
    device=device,
    inputs=input_data,
)
    on_device_output = 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.

Run demo on a cloud-hosted device

You can also run the demo on-device.

python -m qai_hub_models.models.foot_track_net.demo --eval-mode on-device

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.foot_track_net.demo -- --eval-mode on-device

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 Person-Foot-Detection's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

• The license for the original implementation of Person-Foot-Detection can be found here.
• The license for the compiled assets for on-device deployment can be found here

References

• 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.