Stephen Wang Email & Phone Number

Seattle, Wa, Us

San Francisco Bay Area

Working on AI/ML and 3D Vision algorithms in ARKit at Vision Products Group (VPG).

Pittsburgh, Pennsylvania, United States

I built an auto-calibration system and a Structure from Motion (SfM) pipeline to obtain intrinsics efficiently, utilized SuperPoint and Superglue as feature extractor and matcher, and then implemented a learning-based featuremetric refinement inspired by Pixel Perfect SfM to refine 2D keypoints position and 3D triangulated points to improve intrinsic accuracy compared with groundtruth in camera array's KRT.

Pittsburgh, Pennsylvania, United States

I worked on computational imaging under the supervision of Assoc. Prof. Ioannis Gkioulekas, specifically focusing on creating imaging systems that generate feature descriptors and conduct feature matching, and also engaged in research related to physics-based rendering and differentiable rendering.

Belfield, Dublin 4, Ireland

As an undergraduate researcher at @THEIA lab supervised by Dr. Soumyabrata Dev, my research covers various AI-related topics and published 5+ papers. In computer vision, I have worked on salient object detection, stereo matching, 3D reconstruction, and video understanding. In machine learning, I have experience with medical stroke prediction, unsupervised generative models, and computationally efficient ML. As for autonomous driving, I investigated multi-modality and multi-task perception models.2 paper accepted by IEEE ROBIO'231 paper accepted by Elsevier Displays (SCI, IF=4.3)1 paper accepted by Elsevier Healthcare Analytics, with 50+ citations in half-year1 paper accepted by Elsevier Systems and Soft Computing1 paper accepted by Elsevier Entertainment Computing (SCI, IF=2.8)1 work accepted by CVPR'22 Image Matching Challenge (IMC) Workshop1 paper accepted by IEEE ICIP'21

Belfield, Dublin 4, Ireland

As the Teaching Assistant (TA) for several CS modules at UCD (e.g., COMP3025J Augmented and Virtual Reality, COMP3023J Wireless Sensor Networks, COMP2006J Operating Systems), I delivered tutorials and sample exercises to reinforce in-class concepts, and mentored students about lecture contents. In addition, I guided 80+ students in discussion sections, tutored programming assignments, and graded quizzes.

50 Nanyang Avenue, Singapore

I conducted research at the intersection of computer vision, deep learning, and remote sensing with 5+ papers supervised under Assoc Prof Yee Hui Lee at @NTU Energy Research Institute. I proposed a real-time cloud segmentation model that balanced performance and computational complexity, in which I proposed the BSAM module as the decoder to create a segregated feature pair and used Efficientnet-b0 as the backbone to make the model real-time. Finally, the proposed model maintained performance as the SOTA with 70.68% less model size with beyond-real-time-benchmark speed of 299fps and 392fps for FP32 and FP16 respectively.1 paper accepted at IEEE IGARSS (Oral)1 paper accepted by IEEE AP-S/URSI1 paper delivered preprint on arXiv1 paper under review at IEEE TGRS1 paper under review at IEEE ICME

Institute Of Automation, Chinese Academy Of Sciences (Casia)

I developed an on-device real-time object detection module for autonomous driving perception system. As for model building, I integrated MobileNet-YOLOV5 and Faster RCNN, achieving a consensus of predictions. To optimize performance, I implemented k-means++ for calculating adaptive anchor sizes and enhanced the training data with mosaic augmentation. The simulation phase was conducted using CARLA. Furthermore, I utilized Neural Architecture Search to reduce the model size, enabling on-device deployment on NVIDIA Jetson AGX Xavier embedded system-on-module (SoM) for real-time object detection and pose estimation.

Toronto, Ontario, Canada

I was responsible for developing a hybrid deep learning model for MRI image segmentation. I augmented the ABIDE dataset with over 7,8k images using Generative AI models based on Stable Diffusion and used Mip-NeRF for 3D reconstruction. I built a deep learning segmentation model by combining 3D U-Net and Vision Transformer (ViT), enhanced through Boosting ensemble. I also implemented Propensity Score Matching to evaluate time-dependent risk markers for Autism Spectrum Disorder, contributing to early diagnosis efforts.

My task is to build an ML-based recommendation system for the website of instrument sales department to improve monthly revenue. I optimized and fine-tuned several language models like LSTM, Large Language Models (LLM), and BERT to extract/refine the user and product embedding, then combine them into our proposed recommendation deep learning network DeepMDR for personalized recommendation.

Dublin, County Dublin, Ireland

1 paper accepted by IEEE ICIP as first-author1 paper accepted by IEEE ICIVC as first-author (Oral)Thanks to ADAPT SFI Research Centre for long-term support and sponsor for our research work at UCD :)

Master Of Science - Ms, Computer Vision

Grade: 4.11/4.0, Robotics Institute, School of Computer Science.

Master Of Science - Ms, Computer Vision

Bachelor Of Science - Bs, Computer Software Engineering, 3.98/4.0 (Scholaro Algorithm), First Class Honours, Cum Laude

• Relevant Courses: Computer Graphics (A+), Machine Learning (A+), Cloud Computing(A), Augmented and Virtual Reality (A+), Data Structure.