2D Hand-Keypoint-Detection

• This project aim to detect hand joints. Model of this project utilize simple CNN architecture.

Dataset

• Dataset Link: http://www.rovit.ua.es/dataset/mhpdataset/

Project Steps

• Dataset will be consisting of images and 3D location coordinate files of the joints. You need to convert from 3D coordinates to 2D coordinates. Luckily, dataset publishers provided a conversion script. Script is named: 'generate2Dpoints.py'. You need to run this script and generate txt files for the 2D coordinates.

• Then, you need to generate the bounding box coordinate txt files. The relevant script for this task is named 'generateBBoxes.py'.

Note: 'generateBBoxes.py' and 'generate2Dpoints.py' scripts must run with python2. Because this script is written according to python2.

• Okay now.We have hand images,2D coordinates of the joints of the hand images and bounding box of the hand images. Let's get start to preprocess step.

Preprocess Step

• We crop image according to bounding box of the image and padding cropped image to fixed size. Because all images has different cropped size.
• Then joints of the hand images have to readjust according to cropped image.
• All images, joints and bounding box information save into dictionary in order to utilize in model step.
• These steps are implemented in the 'Data Preprocess.ipynb'

Model Step

• Read dictionary that we save. Create CNN model that you want. Set test,train and validation dataset. Fit model and predict over test dataset.
• These steps are implemented in the 'Model.ipynb'

Result

Prediction Images                       Ground Truth Images