Ryan Szeto Email & Phone Number

Remote

Developing computer vision models for improving care for patients with dementia in assisted living facilities.

Remote

Driving Simulation Research: Wrote scenarios for the open-source driving simulator CARLA, and collaborated with research scientists to design experiments for collecting realistic data in simulation.

Ann Arbor, Mi

Projects:* Video Inpainting Benchmark: We developed a benchmark to diagnose important failure modes of video inpainting methods at a large scale. We collected videos and occlusion masks stratified by certain attributes of their content, and we evaluated visual quality via realism, temporal consistency, and reconstruction performance measures. Our experiments demonstrated that inpainting models react to input attributes in ways that trace back to their specific design, highlighting the importance of diagnostic evaluation.* Video Frame Inpainting: We proposed a deep learning solution to video frame inpainting composed of a bidirectional prediction module and a temporally-aware frame blending module. Our approach produces more accurate and qualitatively satisfying results than prior techniques in video inpainting, frame interpolation, and video prediction.* Viewpoint Estimation with Human Guidance: Our Click-Here CNN estimates the location of the camera relative to an object by using a 2D image and the location of a semantic keypoint on the image, provided by a human, as guidance. Our evaluation demonstrated that keypoint information can help models obtain better viewpoint estimates than those that only use information from the image.PI: Prof. Jason J. Corso

San Diego, California, United States

Hyperconsistency for Video-to-Video Translation Tasks: We proposed a solution called hyperconsistency (HyperCon) to reduce the flickering effects that result from applying image processing to video frames independently. To accomplish this goal, HyperCon identifies consistencies in high frame rate space and propagates them into the post-processed result. Our method reduces flickering effects more reliably than prior video consistency work for tasks as disparate as video inpainting and style transfer.Mentor. Dr. Mostafa El-Khamy

Cambridge, Massachusetts, United States

Evaluating the Generality of Video Prediction Models: We investigated the robustness of existing video prediction models to unseen objects and rates of motion. Our experiments revealed that during prediction, certain models reconstruct objects seen during training at inference time, and all that tested models fail to generalize to faster rates of motion.Mentor: Dr. Simon Stent

Littleton, Ma

* Designed and implemented an interface to manage permissions through an access manager* Reduced page loading times by 80% by consolidating remote database queries* Resolved an issue that created detached Java VMs and resulted in out-of-memory errors

Amherst, Ma

PAOL is a system that captures all relevant components of a lecture with inexpensive devices, regardless of the lecturer’s style of teaching. Video frames of the whiteboard and computer screen are processed in close to real time to remove the lecturer and improve the legibility of the content.Responsibilities:* Monitored captured content and processing output for correctness* Modified capture code for compatibility with processing code and web viewer* Developed a laptop-based lecture capture system that can be quickly set up and broken down in any classroom* Investigated methods of improving the clarity of content from whiteboard images* Presented results at various venues, including a Research Experience for Undergraduates (REU) poster presentation session and the 2014 ASEE Zone 1 Conference

Amherst, Ma

The CAD Data Exchange system allows users to convert PTC Creo models to SolidWorks and edit the output, which is not possible through popular CAD data standards. The system implements a theoretical approach that defines a formal grammar for models of each CAD system, as well as translations between the languages. Responsibilities:* Wrote procedures to translate six possible model features for a CAD model translation program* Presented results in a summer Research Experience for Undergraduates (REU) poster presentation

Natick, Ma

* Converted JavaScript unit tests from the Dojo Objective Harness framework to QUnit* Added cut, copy, and paste functionality to the Variable Editor for MATLAB Online

Doctor Of Philosophy (Ph.D.), Computer Science And Engineering

Advisors: Prof. Jason J. Corso, Prof. Honglak Lee

Master Of Science (M.S.), Computer Science And Engineering

Bachelor Of Science (B.S.), Computer Science And Mathematics

Activities and Societies: Phi Beta Kappa, University Chorale* Graduated summa cum laude. * Graduated from Commonwealth Honors College.