Jason Ku

About

As an AI reseacher, my main interests are in Robotics, Computer Vision, and 3D Perception. I recently completed an MASc. at the University of Toronto, researching 3D scene perception for autonomous driving.
2017-2019: MASc. Perception Researcher with Professor Steven L. Waslander
2010-2015: Bachelor's Degree in Mechatronics Engineering at the University of Waterloo
Experienced with Python, C++, Tensorflow, PyTorch, ROS, Git, DVC.

News

2020: MASc completed, papers accepted into ICRA, IROS, and CRV 2020.
Jul. 2019: "Improving 3D Object Detection for Pedestrians with Virtual Multi-View Synthesis Orientation Estimation" was accepted into IROS 2019.
Mar. 2019: Our paper, "Monocular 3D Object Detection Leveraging Accurate Proposals and Shape Reconstruction", was accepted into CVPR 2019. This work was done with Alex Pon and Professor Steven Waslander.
Oct. 2018: Presented the paper, "Joint 3D Proposal Generation and Object Detection from View Aggregation", at IROS 2018. This work was done with Melissa Mozifian, Jungwook Lee, Ali Harakeh, and Professor Steven Waslander.
Sep. 2018: Transferred to the University of Toronto, Institute for Aerospace Studies.

Publications

Improving 3D Object Detection for Pedestrians with Virtual Multi-View Synthesis Orientation Estimation
Jason Ku, Alex D. Pon, Sean Walsh, Steven L. Waslander
IROS 2019: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems
Realistic multi-view synthesis of virtual views of pedestrians to improve heading estimation.
1st place KITTI Pedestrian Orientation Estimation Benchmark
arXiv
Demo Video

Monocular 3D Object Detection Leveraging Accurate Proposals and Shape Reconstruction
Jason Ku*, Alex D. Pon*, Steven L. Waslander (*equal contribution)
CVPR 2019: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition.
Monocular 3D detection network that leverages proposal regression, and is designed to optimize consistency between 2D observations and 3D point cloud estimates instead of 3D bounding box estimates
Demo Video
arXiv / Paper

Joint 3D Proposal Generation and Object Detection from View Aggregation
Jason Ku, Melissa Mozifian, Jungwook Lee, Ali Harakeh, Steven L. Waslander
IROS 2018: 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems.
Open Source, Real-Time 3D Object Detection for Autonomous Driving
Integrated into the autonomoose perception stack and used to drive 100km autonomously
Code / arXiv / Paper
Demo Video
KITTI Benchmarks: 3D / BEV

Other Projects

3D Scene Visualizer
Open source visualization tools for Python using VTK with support for KITTI, nuScenes, and Argoverse
GitHub

Other Experience

3D EEG node position scanning using a handheld tracking camera
The 3D position of up to 255 EEG nodes can be quickly and accurately scanned
SLAM system: camera is localized using visible markers (yellow), new possible markers are processed (red)
Improved robustness of tracking and formulated effective method to remove "phantom" markers
The filtered set of markers (white) represents the best estimated position of the node on the head
Scanned markers are compared to a template and automatically numbered
The entire process can be completed in under 2 minutes

AI Navigation: Replaced way-points with way-gates to improve group cohesion during movement
AI Behaviour: Reworking the perception system to better prioritize agent perceptions based on player and event proximity, and other factors
New Features: "Melee Assist" featuring an auto-target locking system and customizable camera controls to improve the experience for PS4 and PC players using gamepads.
Misc: Gameplay improvements and the release of Steam achievements