Skip to main content

PoseNet MobileNet 0.75 Quant

A quantized PoseNet model using MobileNet architecture with 75% width multiplier for efficient and optimized body pose estimation.

Supported processors

  • SL1680
  • SL1640

Performance

SyNAP

2.32ms
Inference Time
1.84ms
Init Time

TFLite CPU

28.6ms
Inference Time
14.34ms
Init Time
4
Number of Threads

TFLite NPU

6.01ms
Inference Time
381.88ms
Init Time

PoseNet MobileNet 0.75 Quant

Model Overview​

A quantized PoseNet model using MobileNet architecture with 75% width multiplier for efficient and optimized body pose estimation.

The PoseNet MobileNet 0.75 Quant model is developed and optimized for the Synaptics Astra™ SL1680 processor NPU and SL1640 processor NPU.

Model Features​

  • Model Type: Pose Detection
  • Input Size: various resolutions based on deployment
  • Output Size: various resolutions based on deployment

ℹ️ INFO: This model is ready to use on Synaptics Astra Machina boards. An NPU optimized version of the PoseNet MobileNet 0.75 Quant is installed in the Astra SDK Image.

Deployment on Synaptics Astra SL1600 Series​

This particular model is compiled for Synaptics Astra SL1680 processor. You can find this model already pre-installed on Machina™ Dev kit with SL1680 processor.

You can also find the same model compiled for Synaptics Astra SL1640 processor pre-installed on Machina™ Dev kit with SL1640 processor.

Synaptics Astra Machina™ is Modular developer kit for Astra SL-Series of high-performance IoT processors with integrated Synaptics Veros™ wireless connectivity solution. Learn more here

Application binary​

The synap_cli_od command line application allows running object detection models like PoseNet MobileNet 0.75 Quant.

Inputs:

  • The converted synap model (.synap extension)
  • Optionally, a confidence threshold for detected objects
  • One or more images (jpeg or png format)

Outputs:

  • A list of detected objects for each input image, including:
    • Bounding box
    • Class index
    • Confidence score

Command line usage on Astra SL1680 and SL1640:

MODELS=/usr/share/synap/models/

cd $MODELS/object_detection/body_pose/model/posenet_mobilenet_075/posenet_mobilenet_075_quant

synap_cli_od -m model.synap input_image.jpg

Example output on SL1680:

Input image: input_image.jpg (w = 640, h = 480, c = 3)
Detection time: 2.32 ms
#   Score  Class  Position  Size     Description
0   0.95       0   94,193    62,143  person

ℹ️ INFO: JPEG/PNG input images are resized in software to the network input tensor size.

💡NOTE: Ensure the output format is defined during model conversion. Missing format details can result in errors such as "Failed to initialize detector".

Performance on NPU​

ProcessorsInference Time (ms)
SL16802.32
SL16404.13

Optimize and Customize the model​

Advanced users may wish to customize the source model and recompile it for the Synaptics Astra NPU. Please refer to the Bring Your Own Model section for more information.

License​

The source model is licensed under Apache License 2.0.

The compiled model for on-device deployment is covered under the Synaptics Astra EULA.

Learn More​

Related models

Pose Detection

Person Pose Detection 480x640 (SR100 Series)

A Person Pose Detection 480x640 model, developed by Synaptics, is a lightweight quantized `tflite` model developed for the SR100 processors.

Pose Detection

Person Pose Detection 256x480 (SR100 Series)

A Person Pose Detection 256x480 model, developed by Synaptics, is a lightweight quantized `tflite` model developed for the SR100 processors.

Pose Detection

YOLOv8s Pose

A YOLOv8s model specialized for body pose estimation, optimized for real-time applications.

Pose Detection

PoseNet MobileNet 0.75 Float

A PoseNet model using MobileNet architecture with 75% width multiplier for efficient body pose estimation.