Tiancheng Zhao

Building Animatable 3D Model With RGBD Videos

The target is to build animatable 3D Model for deformable objects with RGBD videos casually captured by phones. By extracting segmentation masks of objects, registering dense object features to object meshes and utilizing optical flows to constrain object shape reconstruction and deformation, decent 3D models are created.

Image Editing With Multimodal Instructions

The target is to edit real images with instructions including text, strokes, sketches and masks. Built on a pre-trained Stable Diffusion model, complex editing on content and structure of images is achieved.

Transformer Based Multi-Object Tracking

The target is to track multiple objects with transformers. By modifying DETR architecture, superior performance is achieved and traditional post processing (like NMS, object association) is eliminated.

Multimodal Vehicle Trajectory Prediction
The target is to predict vehicle trajectories with Visual, Lidar, and Map modalities. 3 modalities were adopted collectively to offer enhanced trajectory predictions.
Graph-based Human Action Recognition
The target is to identify actions with deep residual graph neural networks and transformers. Different attention mechanisms and data augmentation methods are evaluated.
Real-Time Hand Gesture Recognition
The target is to detect and classify hand gestures with video cameras in real-time to enable wireless control over computers. The functionalities of the final product include swipe left and right, swift up and down, zoom in and zoom out, turn up (down) the volume, pause (start) the video.

J Yoo, T Zhao, L Akoglu, Data Augmentation is a Hyperparameter: Cherry-picked Self-Supervision for Unsupervised Anomaly Detection is Creating the Illusion of Success, TMLR, 2023