Bicheng Xu
Curriculum Vitae


Bicheng Xu is a Ph.D. student at the University of British Columbia under the supervision of Prof. Leonid Sigal. Previously, he received a Master of Science degree in Computer Science from the University of British Columbia, a Bachlor of Engineering degree in Computer Science and Technology from Zhejiang University, and a Bachlor of Science degree (with distinction) in Applied Sciences from Simon Fraser University with major in Computing Science and minor in Mathematics. His research interests are in Computer Vision, its intersection with Natural Language Processing, and Machine Learning.


- Consistent Multiple Sequence Decoding    [pdf]
  B. Xu, L. Sigal.
  arXiv preprint arXiv:2004.00760, 2020.

- Watch, Listen and Tell: Multi-modal Weakly Supervised Dense Event Captioning    [pdf]
  T. Rahman, B. Xu, L. Sigal.
  IEEE/CVF International Conference on Computer Vision (ICCV), 2019.

- Time Perception Machine: Temporal Point Processes for the When, Where and What of Activity Prediction    [pdf]
  Y. Zhong, B. Xu, G.-T. Zhou, L. Bornn, G. Mori.
  arXiv preprint arXiv:1808.04063, 2018.

- WiLocator: WiFi-sensing Based Real-time Bus Tracking and Arrival Time Prediction in Urban Environments    [pdf]
  W. Liu, J. Liu, H. Jiang, B. Xu,┬áH. Lin, G. Jiang, and J. Xing.
  IEEE International Conference on Distributed Computing Systems (ICDCS), 2016.

Selected Projects

Lyric Generation with Style    [pdf]

  • Built a GAN-like neural network model to generate lyric given a style and a topic
  • Proposed a novel hierarchical structure for both lyric generation and encoding
  • Presented two different evaluation methods to quantitatively measure the authenticity of the generated lyric

Semi-supervised Image Captioning via Reconstruction    [pdf]

  • Proposed an end-to-end model that can generate image captions in a semi-supervised way
  • Adopted the idea of reconstruction to utilize images without paired captions
  • Applied the Gumbel-Softmax approximation to backpropage gradients through word samples during training

Handwritten Chinese Character Generation via Conditional Neural Generative Models    [pdf]

  • Exploited generative adversarial networks (GAN), variational auto-encoders (VAE), and their combinations to generate handwritten Chinese characters conditioned on their GBK encodings
  • Used PyTorch library to build the neural network models

Evaluating Visual Perception with Bouncing Motion    [pdf]  [code]

  • Designed a perceptual experiment and developed a novel interactive interface that supports to investigate human perception to bouncing motions
  • Displayed a rigid ball falling and bouncing from a hill to detect a just noticeable range of plausible motions that loosely follow the law of physics
  • Validated the experiment design and the interface in piloting studies, providing insights to the theory of perception and physical simulation

Connectionist Temporal Classification for Group Activity Recognition in Videos    [pdf]  [poster]

  • Combined VGG Net, recurrent neural network, and connectionist temporal classification (CTC) to recognize a sequence of activities performed by a group of people in a video through supervised learning
  • Constructed a new volleyball dataset using the volleyball game videos available on YouTube.

Distributed File System with Transactional Semantics    [code]

  • Implemented a distributed file system with transactional semantics following the properties of Atomicity, Consistency, Isolation, and Durability using Java
  • Handled omission, byzantine and failstop failures on the client and failstop failures on the server
  • Detail specification