Simulated Action YGAR Hard Level

Our video action dataset is generated using a 3D simulation program developed in Unity. Each data sample consists of a video capturing a human performing various actions. Our initial set of actions comprises a total of 10 different yoga poses: camel, chair, child's pose, lord of the dance, lotus, thunderbolt, triangle, upward dog, warrior II, and warrior III. Within each of these 10 yoga poses, there are four variations, some exhibiting more pronounced differences than others. This results in a total of 40 action types within our dataset.

We opted for this set of actions as our initial dataset because each pose is distinct and has a well-defined ending position. This characteristic allows us to employ image classification techniques to compare both traditional modeling architectures and deep learning approaches.