I’m a huge fan of using synthetic means to create data to drive Machine Learning understanding forward. For computer vision, one often has spend a lot of money to create a dataset of images, by, for example, hiring photographers, models, equipment and the rest.
But with Computer Generated Imagery (CGI), it’s becoming easier to create photorealistic images on your desktop. I started this experiment with the Rock Paper Scissors and Horse or Human datasets, which you can find on this site here
My newest creation is a Yoga Poses dataset, which you can download in its entirety from here
It contains several hundred 300x300 full color images in 5 different Yoga Poses: Downward Dog, The Warrior, The Tree, The Chair, and The Cobra. There are many different backgrounds and both male and female models with a variety of skin and hair tones, as well as different camera angles and lighting. It’s a challenging one to build a classifier with, and hopefully will be a useful alternative to the usual run-of-the-mill datasets!
It also includes many frames of the model going into and out of the pose, instead of just dealing with the finished pose, so you can build a classifier to potentially classify different stages should you want to!
Here’s an example of a male model on his way into the Warrior pose:
Or, the same model a little later into the pose!
Here also is a girl entering the chair pose:
…and the same girl deeper into the pose, from a different angle:
There’s a really cool notebook using it with Movenet – teaching you how to integrate movenet classifications with poses on the TensorFlow site here
…and of course, you could try this dataset with a Convolutional Neural Network, like the ones I teach about on Coursera
I’d love to hear if you find this dataset useful, and please reach out to me on Twitter if so!
All of these images were created using Daz3D should you want to try for yourself!