About

I’m Zheng (Thomas) Tang, a Senior Deep Learning Engineer in NVIDIA’s Cosmos Engineering team. I work on multimodal models, vision-language post-training, and evaluation, spanning data selection, training, benchmarking, and deployment.

My recent work includes:

• Video search and data-efficient learning. Developed multimodal search for the VSS Blueprint, Cosmos3 VLM post-training workflows, and DEFT. The DEFT pipeline exceeded full-data retrieval quality with 37.5% of the training data and 64% less training time.
• Physical AI evaluation. Lead Cosmos physics and geometry benchmarking and drive VLM counting evaluation. Led PhysicalAI-SmartSpaces (PAISS) release and evaluation (900K+ Hugging Face downloads) and contributed evaluation infrastructure for VANTAGE-Bench and Traffic Anomaly Reasoning (TAR).
• Multi-camera 3D perception. Tech-led perception development for the Mega Omniverse Blueprint, demonstrated at CES and GTC 2025.

Previously, I was an Applied Scientist on Amazon One (2019–2021), developing biometric recognition methods. I earned my Ph.D. in Electrical & Computer Engineering from the University of Washington in 2019.

I have led the AI City Challenge organizing committee at CVPR, ICCV, and ECCV since 2020. Across ten editions, the challenge has attracted approximately 2,500 team registrations from 48 countries and regions. I also serve as a Senior Area Editor for IEEE Transactions on Circuits and Systems for Video Technology (T-CSVT).

News

• [Sep. 2026] The Traffic Anomaly Reasoning (TAR) and PhysicalAI-SmartSpaces (PAISS) benchmark papers were accepted to NeurIPS 2026
• [Sep. 2026] Successfully concluded the 10th AI City Challenge at ECCV 2026, with 2.5K teams across 48 countries and regions engaged over the decade-long challenge, the 2026 results featured by NVIDIA AI, ECCV Daily News, and an iMerit interview
• [Jun. 2026] Supported VANTAGE-Bench and TAR benchmark efforts highlighted with NVIDIA Cosmos 3 at Computex 2026, including the VANTAGE-Bench evaluation system release and TAR launch alignment for AI City Challenge 2026 Track 3 at ECCV 2026
• [Apr. 2026] The 10th AI City Challenge was accepted as a half-day workshop at ECCV 2026 in Malmo, advancing Sim2Real transfer, unified reasoning, and smart-city video AI across six tracks
• [Mar. 2026] Supported the GTC 2026 VSS booth demo for video analytics AI agents, contributing to video embedding search, search-model preparation, demo evaluation data, and VSS 3.1.0 search components
• [Dec. 2025] Presented NVIDIA Metropolis multi-camera 3D perception and spatial AI research at the UrbanAI Workshop at NeurIPS 2025, covering AI City Challenge progress, cloud-native tracking workflows, MCBLT, and Sparse4D
• [Oct. 2025] Vision AI agents for Foxconn factory digital twins showcased in NVIDIA GTC DC 2025 Keynote, and hosted the 9th AI City Challenge Workshop at ICCV 2025 in Honolulu with 245 teams worldwide
• [Mar. 2025] Metropolis spatial AI agents, integrated into the Mega Omniverse Blueprint, featured in CES 2025 & NVIDIA GTC 2025 Keynotes by CEO Jensen Huang
• [Mar. 2025] Released MTMC (multi-target multi-camera) tracking datasets for AI City Challenge 2024 & 2025 as part of the Open Physical AI Dataset at NVIDIA GTC 2025
• [Jan. 2025] Won Best Presentation and Best Poster Awards in the Imaging and Computer Vision Track at NTECH 2024 for BEV-SUSHI and MTMC-to-RTLS work
• [Jan. 2025] Promoted to Senior Area Editor (SAE) of IEEE T-CSVT
• [Oct. 2024] Served as Area Chair for ACM Multimedia 2024
• [Jul. 2024] Paper titled "UAV First-Person Viewers Are Radiance Field Learners" accepted to ECCV 2024
• [Jun. 2024] Creation of the largest indoor synthetic dataset to support the 8th AI City Challenge at CVPR 2024 featured in an official NVIDIA blog
• [Mar. 2024] Fusing Real-Time AI With Digital Twins demo highlighted in NVIDIA GTC 2024 Keynote by CEO Jensen Huang
• [Mar. 2023] Released NVIDIA Metropolis Microservices 1.0, including the Multi-Camera Tracking app I developed and trained using synthetic Omniverse data
• [Jan. 2023] Released the 7th AI City Challenge Workshop in conjunction with CVPR 2023
• [Dec. 2022] Released NVIDIA TAO Toolkit 4.0, where I developed re-identification and pose-based action recognition networks, and end-to-end video analytics pipelines
• [Oct. 2022] Special issue "Label-Efficient Learning on Video Data" accepted to T-CSVT
• [Feb. 2022] Received the 2021 Best Associate Editor Award from IEEE T-CSVT
• [Jan. 2022] Released the 6th AI City Challenge Workshop in conjunction with CVPR 2022
• [Oct. 2021] Served as Area Chair for MLSP 2021
• [May 2021] Joined the NVIDIA Metropolis team as a Senior Data Science Engineer
• [Jan. 2021] Released the 5th AI City Challenge Workshop in conjunction with CVPR 2021
• [Dec. 2020] Appointed Associate Editor for IEEE T-CSVT (Impact Factor: 5.859)
• [Sep. 2020] Amazon One launched; contributed as a research scientist and filed two U.S. patents
• [May 2020] Source code for “PAMTRI: Pose-Aware Multi-Task Learning for Vehicle Re-Identification Using Highly Randomized Synthetic Data” (ICCV 2019) released on GitHub
• [Jan. 2020] Released the 4th AI City Challenge Workshop in conjunction with CVPR 2020
• [Jul. 2019] “PAMTRI” paper accepted to ICCV 2019
• [Jun. 2019] Earned Ph.D. in Electrical & Computer Engineering from University of Washington
• [Feb. 2019] “CityFlow” benchmark paper accepted to CVPR 2019 (Oral)
• [Jun. 2018] Led the #1 winning team in Track 1: Traffic Flow Analysis (Demo) and Track 3: Multi-Camera Vehicle Detection and Re-ID (Demo) at the 2nd AI City Challenge Workshop, CVPR 2018
• [Jun. 2018] “Joint Multi-View People Tracking and Pose Estimation for 3D Scene Reconstruction” accepted to ICME 2018 (Oral)
• [Aug. 2017] Winner of Track 2: AI City Applications (Demo) at the 1st AI City Challenge Workshop, SmartWorld 2017
• [May 2017] “Online-Learning-Based Human Tracking Across Non-Overlapping Cameras” accepted to T-CSVT
• [Dec. 2016] Finalist for two Best Student Paper Awards at ICPR 2016 for “Camera Self-Calibration from Tracking of Moving Persons”
• [Dec. 2015] “Multiple-Kernel Adaptive Segmentation and Tracking (MAST)” accepted to ICASSP 2016 (Oral)
• [Jun. 2014] Graduated with First Class Honours from the BUPT-QMUL Joint Programme (B.Sc. in Engineering)