[1I5-GS-2] Machine learning: Applied machine learning (2)

[1I5-GS-2-04] Unsupervised Motion Feature Extraction in Video via Bi-directional GAN

〇Yuma Uchiumi¹, Yuki Abe¹, Takuma Seno¹, Michita Imai¹ (1. Keio University)

Keywords:Machine Learning, Video Generation, Generative Adversarial Network , Deep Learning

When extracting a certain motion feature of the object from video data,
it is necessary to capture not the consistent pattern for the object recognition
but the sequential pattern for the motion recognition.
To handle this problem, we propose Motion Disentangled Bidirectional GAN (MDBiGAN) that
disentangles the latent variable of Bidirectional GAN.
Then, we show that MDBiGAN clearly extracts the motion features from video data as the experimental result.