Sitemap

A list of all the posts and pages found on the site. For you robots out there is an XML version available for digesting as well.

Pages

About me

I am currently a Ph.D. student in the University of Illinois Urbana-Champaign (UIUC), advised by Alexander Schwing. Previously, I obtained a MPhil degree from The Hong Kong University of Science and Technology (HKUST)1, advised by Yu-Wing Tai and Chi-Keung Tang.

I am recently working on scalable (long-term and open-world) video segmentation and understanding. I am generally interested in artificial intelligence. I believe in AGIs and has high hope for their potential to transform human civilization for the better.

I am delighted to have worked with Seoung Wug Oh, Brian Price, and Joon-Young Lee at Adobe Research in 2022.

Name

“Ho Kei” (with the space) is my first name and “Cheng” is my last name. “Rex” is the commonly used “english name” that is not part of my legal name.

News:

  • (Jul 2022) XMem is accepted to ECCV 2022.
  • (Sep 2021) STCN is accepted to NeurIPS 2021.
  • (May 2021) In the YouTubeVOS 2021 challenge, STCN achieved 1st place accuracy in novel (unknown) classes and 2nd place in overall accuracy.
  • (Mar 2021) MiVOS is accepted to CVPR 2021.
  • (Mar 2020) CascadePSP is accepted to CVPR 2020.
  1. Yes, “The” should always be capitalized for HKUST. This is not the case for UIUC. I also received my bachelor degree from HKUST. 











misc

Warcraft III Map Making

I got my first taste of programming in secondary school, making maps (similar to mods) for a game Warcraft III. It started as an anti-cheat patch for an existing map (兵戰), and I later started adding new features. At some point, I completely reprogrammed the entire map. The tradeoff between game balancing and fun was the most daunting task.

18 versions were released from 2011 to 2016.

war3


Cats @ HKUST

drawing drawing drawing
drawing drawing drawing
drawing drawing drawing


Robotics Team

I joined the HKUST Robotics Team in UG year 1. We missed champion in Hong Kong Robocon 2016 with the lack of experience and struck back to win champion in 2017. We represented Hong Kong in the International Robocon 2017 competition in Tokyo and took home the Best Design Award.

I developed various low-level libraries (communication protocol, PID, peripheral sensors) and high-level functionalities (path planning, state machines), and picked up cruical engineering/problem solving skills in this journey.

I was the chairperson for the 8th Robot Design Contest in HKUST for new member recruitment. Grateful to this amazing team and all the enthusiastic teammates!

Credit: Radio Television Hong Kong (RTHK).


publications

CascadePSP: Toward Class-Agnostic and Very High-Resolution Segmentation via Global and Local Refinement

CascadePSP: Toward Class-Agnostic and Very High-Resolution Segmentation via Global and Local Refinement

Ho Kei Cheng*, Jihoon Chung*, Yu-Wing Tai, Chi-Keung Tang. Published in CVPR, 2020.

Refines segmentations (4K and beyond) in a class-agnostic manner without using any high-resolution training data through a set of carefully designed cascade operations.


Modular Interactive Video Object Segmentation: Interaction-to-Mask, Propagation and Difference-Aware Fusion

Modular Interactive Video Object Segmentation: Interaction-to-Mask, Propagation and Difference-Aware Fusion

Ho Kei Cheng, Yu-Wing Tai, Chi-Keung Tang. Published in CVPR, 2021.

Used by: [Sieve], [Trioscope]

A more user-friendly and efficient paradigm of iVOS in which interactions and propagations are decoupled, with the user’s intention captured by a novel difference-aware fusion module.


XMem: Long-Term Video Object Segmentation with an Atkinson-Shiffrin Memory Model

XMem: Long-Term Video Object Segmentation with an Atkinson-Shiffrin Memory Model

Ho Kei Cheng, Alexander Schwing. Published in ECCV, 2022.
Also presented in Computer Vision for Metaverse Workshop 2022, Workshop on Computer Vision in the Wild , Workshop on AI for Creative Video Editing and Understanding, and In-vehicle Sensing and Monitorization Workshop.

We develop a multi-store memory model to untie accuracy with memory consumption -- achieving good results in both short and long videos.