Aloha my name is Ringo and welcome to my page! I am now a researcher at SenseTime Research. Previously I completed my undergrad degree in Computer Science at University College London. My research lies on scene understanding (e.g image correspondence/visual tracking/relation), and embedded ML systems. I was a research intern with Intel and Dyson, where I worked on robotic and more vision projects. Prior to that, I was also a renowned researcher at Imperial College London Custom Computing, where I worked on NASA AVRIS imaging algorithm and inference acceleration with FPGA.

Apart from being a nerd, I am a bookworm, a tennis player. I also collect cassette tapes. If you want to know more about me, you can look at my Google Scholar, Facebook, LinkedIn, my resume or even contact me in person. :)

Publication and Projects

Below is the collection of my researches and some of my interesting projects! (That can be displayed in the public)

Epipolar Shifted Rectangular Window Transformer for Stereo Matching

Ringo S.W. Chu, Jinjin Gu, Jimmy S. Ren
Under Review
[Paper] [Code] [Project Page]

Abstract: Upcoming...

Self-Supervised Intensity-Event Stereo Matching

Jinjin Gu, Jinan Zhou, Ringo S.W. Chu, Yan Chen, Jiawei Zhang, Xuanye Cheng, Song Zhang, Jimmy S. Ren
Journal of Imaging Science & Technology, 2023
[Paper] [Code] [Project Page]

Abstract: One of the most foremost system for handling stereo matching from different modularity.

Visual Relation Learning through Transformer Network

A Project with LUKADVISOR investigating the use of Transformer to solve Visual Relation Detection. VRD was tackled by constructing graphs or guided by language models. Motivated by the recent success of Transformer in linguistic and vision tasks, this project aims to use the transformer architecture as a Look-Up-Table to directly predict the predicate(relation) between subject-object pairs.

(A Closer Look at) Neural RGB+𝔻 Object Tracking with Depth Estimation

This project introduces the first neural approach to support 2D tracking using approximated depth information. We assume that objects should have small displacement and small change in depth displacement between frames(constant velocity assumption), and apply this assumption to a visual object tracker and a multiple object tracker using depth information generated from a monocular depth estimation network. Based on preliminary quantitative results, the proposed approach has shown promise in occlusion detection for tracking of single and multiple generic objects.

Quantization Sim2Real Robotic Arm Control

Actual data is expensive! Collecting huge number of data, or training on a real device is expensive and super time-consuming. In reality, robotic models are trained on simulation instead. However, one of the problems imposed from simulation is that- Simulation differs from reality, which is often called the reality gaps. One promising results were shown is to add Gaussian noise to your simulation input, but this project is instead stealing information away from the input to achieve a ‘randomness’. This project quantizes the parameter using 16-fix, 32-fix, 16-32 mixed fix of the custom simulation, and perform training on various simulation based on various RL algorithm.

Ho-Cheung Ng, Steven Liu, Izaak Coleman, Ringo Chu, Man-Chung Yue, Wayne Luk
The International Conference on Field-Programmable Technology, 2020
[Paper] [Bibtex]

Abstract: This paper explores the use of a reconfigurable architecture to accelerate this process, based on the seed-and-extend model of Bowtie2. In the proposed approach, complete information available in sequencing data is integrated into an FPGA alignment pipeline for biologically accurate runtime acceleration. Experimental results show that our architecture achieves a similar alignment rate compared to Bowtie2, mapping reads around twice as fast. Particularly, the alignment time is reduced from 50 minutes to 26 minutes when processing 300M reads.

Martin Ferianc, Hongxiang Fan, Ringo S.W. Chu, Jakub Stano, Wayne Luk
Applied Reconfigurable Computing (ARC), 2020
[Paper] [Bibtex]

Abstract: A method for exploring the design space and latency prediction for FPGAs using a Gaussian process parametrized by an analytic approximation and coupled with runtime data. Experiments were conducted on three CNNs accelerator targeting Intel Arria 10 GX 1150, demonstrating a 30.7% improvement in accuracy w.r.t MAE against a standard analytic method in leave-one-out cross-validation.

Ringo Chu, Ho-Cheung Ng, Xiwei Wang, Wayne Luk
IEEE Symposium On Geoscience and Remote Sensing (IGARSS), 2019
[Paper] [Code] [Bibtex]

A new deep-learning model to extract 3D features at spatial and spectral levels simultaneously to address the classic problem of high dimensionality on HSIs. Experiments were conducted using scenes acquired by NASA Airborne Visible/Infra-Red Imaging Spectrometer and have shown competitive performance.

{Steven Liu*, Ringo Chu*}, Xiwei Wang, Wayne Luk
Applied Reconfigurable Computing (ARC), 2019
[Paper] [Code] [Bibtex]

An FPGA accelerator for hyperspectral imagery with an optimized architecture that is faster than GPUs and consume less power. The proposal method is suitable and ideal for real time inference deployment.

Multiple Agents Artificial Stock Market Modelling

[Slides] [Code]

A research project to model a stock market with agents of many behaviours. A probabilistic method is used to analyse the stochasticity of processes and collectives of agents. The model is also benchmarked on Intel Xeon multi-core CPU and Altera FPGA.

LSTM verses Gaussian Process Regression

Me, 2018 Summer
Short Summary

A very very very brief comparison between Long-Short term memory and Gaussian Process Regression on stock predictions based on their interpretabilities.

Mindstorm Robots


My childhood, and then recently

First prototype born in 2015(old photo and rare footage). This is a 35cm x 35cm x 40cm sized robot built in Lego only! The robot was initially constructed with a Sony phone attached as an explorer and as a Mars rover capable of broadcasting live motions. Later I experimented some Reinforcement learning algorithms with the robot for solving maze and learning to escape from being stuck.