Background and Purpose
To improve golf skill, we used to watch videos or invite a coach for teaching. However, it’s hard to learn correct swing forum by watching videos only.And, a expensive training fee is necessary to pay for inviting a golf coach.
This research aims to discover a simple method for correcting golf swing forum, and make it into a usable system. This research is a collaboration research with a certain company.
Introduction
This system is based on Kinect. We use cameras like Kinect, take a photo, and make it into a depth image. Then, input the depth image into a pre-trained model, analyze the swing pose information containing in that image. There is many necessary information for analyze a swing pose, currently we are focusing on extracting joints’ position.
1.Training networks due to CG Images
This research is using deep learning to analyze swing pose. To train a reliable network, data from various person doing random swings is necessary. However, getting those data means that we have to take depth photos and record the joints’ position, which might take a lot of time and cost.
Considering the problem above, this research is using CG images for training the network. The detail is shown in the following figure. We use the swing pose data actually taken by human, then adapt the pose to a CG character with random figure using a 3DCG program called Blender, and take depth photos. Using that method, we can make a variety of depth image using a single pose data, in order to expanding dataset’s scale easily.
2.How to expand dataset’s scale – Swing Pose Data
We expanded our dataset’s scale by importing one swing pose data into multiple CG models which have various figure.
3.How to expand dataset’s scale – Depth Image
We import the expanded swing pose data into multiple CG models, then save the depth image. The following figure shows the combination of 3 types of CG models and one swing pose data. By doing the work above, we can expand our dataset in a simple way.
4.Predicted results and current works
After training the network with dataset for training, we validate the accuracy using dataset for validation. The following video shows the prediction of a certain swing pose plotted on a 3D space(red-line skeleton are ground truth, blue-line skeleton is predicted result). Of course the final purpose is to predict a real image, but currently we are conducting the research about higher accuracy with improving network structure and dataset’s quality.
