David Ross
Research Manager at Google
San Jose, California, United States
Summary
👤
Senior🎓
Top SchoolDavid Ross is a research manager and computer vision leader with 8+ years driving video understanding at Google, now leading Visual Information and Dynamics efforts at Google DeepMind. He built and shipped foundational systems — from the TensorFlow Object Detection API and object tracking in Cloud Video Intelligence to human action recognition in Google Photos — and organized the influential AVA Challenges that advanced spatiotemporal action recognition. Earlier he led YouTube Mix, the personalized radio engine for YouTube Music, and developed large-scale perceptual systems like Melody Match and celebrity face recognition. He holds a Ph.D. in Machine Learning and Computer Vision from the University of Toronto and combines deep academic rigor with product-grade engineering. His open-source contributions include precise, microsecond-aware evaluation work for the ActivityNet/AVA benchmarks, reflecting a focus on measurement fidelity as well as model innovation. Based in San Jose, he blends research publishing at top venues with hands-on system design that powers consumer and cloud-scale video products.
8 years of coding experience
9 years of employment as a software developer
Ph.D., Machine Learning & Computer Visio, Ph.D., Machine Learning & Computer Visio at University of Toronto
Github Skills (6)
computer-vision10
machine-learning10
python10
evaluation10
metric10
numpy10
Programming languages (3)
Github contributions (5)
activitynet/ActivityNet
Apr 2018 - May 2020
This repository is intended to host tools and demos for ActivityNet
Role in this project:
ML Engineer
ML EngineerContributions:6 commits, 4 PRs, 2 comments in 2 years 1 month
Contributions summary:David focused on enhancing the evaluation scripts for the AVA dataset within the ActivityNet repository. Their work involved refining action detection performance calculations, cleaning up logging procedures, and adjusting the handling of warnings and informational messages. Furthermore, the user optimized the AVA evaluation code, removed unnecessary code paths and capped the number of detections per keyframe. They also addressed issues related to distractor frames and added support for microsecond-precision timestamps, which is crucial for precise evaluation.
hostwebapplicationactivitynetweb-applicationactivitypub
cvdfoundation/ava-dataset
Mar 2018 - Mar 2019
The AVA dataset densely annotates 80 atomic visual actions in 351k movie clips with actions localized in space and time, resulting in 1.65M action labels with multiple labels per human occurring frequently.
Contributions:6 commits, 3 pushes, 2 comments in 1 year
avadatasetmovieclipslocalized
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.