Teaser image for the tasks introduced in (i, green box) and (ii, magenta box) in Research Theme1.

We aim to present natural human motion to solve computer vision/graphics tasks. Specifically, we attempt to provide better solutions to human motion prediction/generation/control using deep learning-based techniques. The research of this filed has a great impact on real world applications. For example, (i) Given only a single image as input, our method gives multiple future motion hypotheses to address the two-fold stochasticity imposed by the single image and human motion; (ii) Given an input past motion, our method predicts multiple future motions with user-specified body parts controlled.

Teaser image for the tasks introduced in (i, orange box) and (ii, blue box) in Research Theme2.

We work on topics on image generation/editing with deep learning and pattern recognition techniques. In particular, we aim to develop probabilistic modeling frameworks to manipulate the pixel distributions to edit images. For example. (i) we can modify the color distribution of an image to transfer its color style; (ii) We can recover the latent sharp image from the blurred image.