pinto0309/mobilenetv2-poseestimation

Tensorflow based Fast Pose estimation. OpenVINO, Tensorflow Lite, NCS, NCS2 + Python.

MobileNetV2-PoseEstimation

[Caution] The behavior of RraspberryPi+NCS2 is very unstable.
[Caution] The behavior of Tensorflow Lite+CPU is unstable.
[Caution] May 06, 2019, The Google Edge TPU program and model are under construction.
[Info] Jun 08, 2020, I'm tuning the performance of the Tensorflow Lite model significantly. https://github.com/PINTO0309/PINTO_model_zoo/tree/master/007_mobilenetv2-poseestimation

Introduction

This repository has its own implementation, impressed by ildoonet's achievements.
Thank you, ildoonet.
https://github.com/ildoonet/tf-pose-estimation.git

I will make his implementation even faster with CPU only.

Environment

• Ubuntu 16.04 x86_64
• OpenVINO 2019 R1.0.1
• Tensorflow v1.12.0 + Tensorflow Lite
• USB Camera
• Neural Compute Stick 2 (NCS2)
• Google Edge TPU
• Python 3.5

Environment construction and training procedure

Learn "Openpose" from scratch with MobileNetv2 + MS-COCO and deploy it to OpenVINO/TensorflowLite Part.1

Learn "Openpose" from scratch with MobileNetv2 + MS-COCO and deploy it to OpenVINO/TensorflowLite (Inference by OpenVINO/NCS2) Part.2

Core i7 only + OpenVINO + Openpose Large model + Sync mode (disabled GPU)

NCS2 x1 + OpenVINO + Openpose Large model + Async + Normal mode

Core i7 only + OpenVINO + Openpose Small model + Sync + Boost mode (disabled GPU)

NCS2 x1 + OpenVINO + Openpose Small model + Async + Boost mode

Usage

$ git clone https://github.com/PINTO0309/MobileNetV2-PoseEstimation.git
$ cd MobileNetV2-PoseEstimation

CPU - Sync Mode

$ python3 openvino-usbcamera-cpu-ncs2-sync.py -d CPU

CPU - Sync + Boost Mode

$ python3 openvino-usbcamera-cpu-ncs2-sync.py -d CPU -b True

NCS2 - Sync Mode

$ python3 openvino-usbcamera-cpu-ncs2-sync.py -d MYRIAD

CPU - Async Mode

$ python3 openvino-usbcamera-cpu-ncs2-async.py -d CPU

NCS2 - Async - Single Stick Mode

$ python3 openvino-usbcamera-cpu-ncs2-async.py -d MYRIAD

NCS2 - Async - Multi Stick Mode

$ python3 openvino-usbcamera-cpu-ncs2-async.py -d MYRIAD -numncs 2

NCS2 - Async - Single Stick + Boost Mode

$ python3 openvino-usbcamera-cpu-ncs2-async.py -d MYRIAD -b True

GPU (Intel HD series only) - Async - Boost Mode

$ python3 openvino-usbcamera-cpu-ncs2-async.py -d GPU -b True

Reference articles, Very Thanks!!

https://github.com/ildoonet/tf-pose-estimation.git
https://www.tensorflow.org/api_docs/python/tf/image/resize_area
Python OpenCVの基礎 resieで画像サイズを変えてみる - Pythonの学習の過程とか - ピーハイ
Blurring and Smoothing - OpenCV with Python for Image and Video Analysis 8
https://www.learnopencv.com/deep-learning-based-human-pose-estimation-using-opencv-cpp-python/
https://teratail.com/questions/169393

mobilenetv2-poseestimation's Issues

TFlite Models won't produce any pose, only random points everywhere

The Openvino scripts and models are working pretty good, but the tflite ones (normal tflite and edge tpu) are only producing random points as poses everyhere in the picture, no matter if there are people visible, or not.

I'll attach a Screenshot to show the problem:

OpenVINO 2022.1 not supported

OpenVINO 2022.1 seems not to be supported. Would it be possible to re-convert the files with the newest version?

RuntimeError: The support of IR v5 has been removed from the product. Please, convert the original model using the Model Optimizer which comes with this version of the OpenVINO to generate supported IR version.

I tried to convert the model with the 2022.1 open model optimizer, but it seems that it does not work. I used the frozen model form the checkpoint path and after conversion the bin file is only 2.2MB (v1.4). And the results are completely wrong, it seems there are weights missing.

Could you give me a hint how to convert it for more recent OpenVINO versions?

-

Broken download link for updated quantized model

[models/train/test/tflite/latest/download.sh] link in this file seems to be broken. Please share the link to download model checkpoints or frozen graph for the updated quantized models.

Unsupported primitive of type: Interp name: base/layer_14/output/upsample with the CPU models.

Hi. Great work. Still, I am getting an error with the CPU models on Windows with Openvino. Can you help me with this please? Thanks.

FPS rate for Raspberry pi

I am running NCS2 - Async - Single Stick + Boost Mode with one NCS on the raspberry pi,
but the frame rate of detection is <2FPS.

Run not via openvino

Hi,
Thank for your contribute. I just want to run openpose with tf lite not using openvino, could you give me a direction?