/YOPO

You Only Pose Once - Neural network for pose Estimation

Primary LanguagePythonGNU General Public License v3.0GPL-3.0

Build Status

YOPO

Is an adaptation of the YOLOv1 algorithm and made using Darkflow. Where by a new framework is added that is based on the original YOLOv1 framework in Darkflow.

YOPO is split into two parts: the pre-processing and the YOPO network.

Getting started

  1. Let pip install darkflow globally in dev mode (still globally accessible, but changes to the code immediately take effect)

    pip3 install -e .

  2. Install with pip globally

    pip3 install -r requirements.txt --user

  3. The human pose dataset weights can be downloaded form here - https://1drv.ms/f/s!Ao7rGgBsjbqiiqwNhPhjgikppDSgMw If no weight files are present please contact rij12@aber.ac.uk

YOPO preprocessing

The preprocessing part requires:

  1. The MPII dataset images which can be found here: http://datasets.d2.mpi-inf.mpg.de/andriluka14cvpr/mpii_human_pose_v1.tar.gz
  2. The MatLab data file containing the metadata about the dataset here (default location in YOPO_preprocessing/data): http://datasets.d2.mpi-inf.mpg.de/andriluka14cvpr/mpii_human_pose_v1_u12_2.zip
  3. The config entries to be set your your configuration(default location: YOPO/YOPO_preprocessing/src/main/conig.py), for example set the location of the MPII dataset images and MatLab data file.

YOPO Network

  1. The project was developed with pycharm so to run it from the command you must export the path for example:

    export PYTHONPATH=/home/USER/git/YOPO

  2. Just build the Cython extensions in place.

    python3 setup.py build_ext --inplace

  3. Each limb is defined in a text file labels.txt

  4. The preprocessing must have been completed with images and label folder created(default: YOPO/YOPO_preprocessing/data/darkflow).

  5. Training the network:

    For any addtional options use:

    python3 ./flow --h
    • --model is the network configuration file, that defines that network.
    • --load is used to supply weights to use as a starting point, as discussed in the report YOPO uses the original weights.
    • --annotations The file path with a folder containing the labels for the training images
    • --dataset The file path with a folder containing the images that the network will train on.
    • --epoch How many the epoch should be executed before the network stops.
    • --gpu Optional flag but recommended. The percentage of GPU usage that Darkflow is allows to use for training the network.
    • --summary Optional flag for the TensorBoard event file to be save for later use in TensorBoard.
    • --batch defines the amount of image that be processed through the network at once. If you have a small amount of RAM or VRAM use a low batchsize.
    python3 ./flow --model NETWORK_CONFIG_PATH --load YOLO_WEIGHTS_PATH --train --annotation LABELS_PATH --dataset DATASET_PATH --epoch 20000 --gpu 0.9 --summary OUTPATH --batch 8

The weight outputted from the training of the network are save in a folder called ckpt that will be create during training.

  1. Testing the network

    Where

    • --imgdir is the image that you would like to train on
    • --model is the network configuration file
    • --load is the weight file that was outputted from the network, -1 can be used to load the last outputted weight from the network.
    • --threshold the minimum probability a box must have before it's shown on the output images
    python3 ./flow --imgdir sub_set/images --model /home/richard/git/YOPO/cfg/yopo.cfg --load -1 --threshold 0.1

Run Tests

IOU unit tests:

python3 test/YOPO/test_calculate_IOU.py

Darkflow tests that require "buildPath" to be change to cloned repo root.

pytest -x --cov=./

Acknowledgments

This project is an adapted folk of Darkflow (https://github.com/thtrieu/darkflow), with the attempt of creating a human pose estimation system.

YOLO papers, Real-time object detection and classification. Paper: version 1, version 2.

Read more about YOLO (in darknet) and download weight files here.