Projects

Training the deep learning model to automatically detect the mahjong tiles and choose the best tile for each turn.

A fisheye-aware transformer-based model designed for 3D egocentric human mesh recovery. We propose an egocentric position embedding block to generate an ego-specific position table for the Swin Transformer to reduce fisheye image distortion.

Apply the GAN to generate the layout of the furniture in the bedroom, living room, dining room and bath room.

This is image compression for machine, which propose a transformer-based context entropy model.

This is transformer-based image compression method, which propose a 2D zigzag entropy model. The paper of this project won the Best Student Paper Award.

We utilize the solutions of Monge–Ampère Partial Differential Equation (MAPDE) as the new loss function of WGAN to make the training process more stable.