victimsnino / pose-simple-baselines-demo.pytorch

CV2 and OpenVINO demo code for the paper https://arxiv.org/abs/1804.06208

Demo of Simple Baselines for Human Pose Estimation and Tracking

This repository contains demo code for the paper Simple Baselines for Human Pose Estimation and Tracking.
Original repository (https://github.com/Microsoft/human-pose-estimation.pytorch)
This repository created at Intel R&D lab at HSE, Nizhny Novgorod.

Introduction

This code estimates keypoints of one person, which centered on the image (inside blue rectangle)

Requirements

• Python 3.6
• PyTorch 0.4.1 (should also work with 1.0, but not tested)
• (Optional) Install dependencies from original repository

Webcam Demo

1. Download required models from original repository from step 8 of Installation (for example 256x256 model
2. run script (for this model)
   python demo.py --cfg pose_resnet_50_256x256.pth.tar.yaml

Testing

1. Download required models from original repository from step 8 of Installation (for example 256x256 model
2. Prepare image, that you want to use for testing
3. run script:
   python demo.py --cfg <path to cfg *.yaml> --image-file <path to image> [--save-transform-image] [--skip-crop-mode] [--gpus <id> ] [--min-confidence-threshold <coef> ]

Description of args:

• cfg (only for demo.py) : You should choose config with the same name as a model, that you are want to use. This file includes different configs for these models.
• image-file : Path to your image for predication. You can didn't use this argument, then will be used webcam
• model-file (only for openvino-demo.py) : You should set it to your *.xml model
• save-transform-image: You can set it for saving temp image after resizing and drawing bounding box (it works for webcam too)
• skip-crop-mode : Use crop mode for cropping person, that are you want (after adding this parameter, you get a new window with your photo, where you should highlight a required zone). By default crop mode is on. You can skip it with help this key
• min-confidence-threshold : Minumal confidence threshold of joints, that will be drawing on image. Default: 0.5

Export models for OpenVINO

For using OpenVINO script you shold convert original models into *.onnx and then to *.bin and *.xml. For this you need:

1. Put models like for using in ./models
2. Run export.py with --cfg argument
3. In folder with models you can find new file *.onnx
4. This file you shold insert, like argument for model optimizer of OpenVINO (OPENVINO_ROOT/deployment_tools/model_optimizer/mo.py). After this you get a *.bin and *.xml for openvino demo script

Example of using

This repository includes an example image. example.png:

If we try to run our script with a key --skip-crop-mode, we get an image without or wrong keypoints and message to console "Bad position of person! Can't find key points. Please, place it at the center of the image or use crop mode for this". And get this image with bad predications:

For correct prediction, we should place the person at the center of the image or use crop-mode for selecting him, if our image is not correct. For example, we can select our person like this:

And get the correct result: