Yu Tsz Ho

Yu Tsz Ho

Tsz-Ho YU is currently a Ph.D. student in the Machine Intelligence Laboratory, Department of Engineering (CUED) of the University of Cambridge, supervised by Professor Roberto Cipolla. His research is sponsored by the Croucher Foundation Scholarship and Cambridge Overseas Trust.

Previously, he received the B.Eng. (first-class honours) degree in Computer Engineering from the Chinese University of Hong Kong in 2007, and the M.Phil degree in Computer Science and Engineering from the same school, under the supervision of Prof. Yiu-Sang Moon, in 2009.

Tsz-Ho is a member of Darwin Collge, University of Cambridge. Tsz-Ho's research interests include computer vision, machine learning and image processing. His main area of research involves using machine learning techniques to discover high-level structures for visual objects, particularly for three-dimensional data (e.g. 3D shapes and videos)

Current work

A common theme of my research is "Discovery High-level Structure of Objects for Registration and Recognition".

Many computer vision applications, such as object recognition and segmentation, are high-level problems that involve understanding the semantic of input data. However, they are often approached by learning statistics on low-level features, such as texture patches or optical flow, bypassing the high-level shape information. Low-level features become less reliable than high-level structures in tasks such as 3D shape recognition and image classification, where texture features are absent or having a weak discriminative power. Dicovering the high-level structure of objects is essential for answering many fundamental computer vision problems, such as 3D object recognition and pose estimation.

The visual representation of an object is not limited to typical two-dimensional images and textures, objects can also be represented visually using videos, three-dimensional shapes and sensor data (e.g. depth maps from range scanners). These data usually provide extra information about an object, such as its structure and motion/action of the object. Utilzing these new types of information, it is possible to extract more reliable features (hints) for recongizing an object, or analyzing its attributes. It is both interesting and useful to develop new approaches for recognizing things from 3D shapes and videos, using machine learning techniques.

The main focus of my current research project is using structural information for object registration and recognition. I have been developing both discriminative (e.g. Random Forest) and generative models (e.g. Constellation model) to discover object structure for both registration and recognition.