The lab is proud to announce that Michael Perez has successfully defended his PhD dissertation, “Toward Deployable Action Understanding for Real-Time Intelligent Systems: System Integration, Benchmarking, and Efficient Temporal Modeling.” Congratulations, Dr. Perez!
Michael’s research tackles a central challenge in computer vision: how to understand human activity from long, continuous video in real time, without relying on massive labeled datasets or heavy computational budgets. His dissertation makes four contributions toward this goal. ENKIx, an augmented reality task guidance system, combines object detection, action recognition, task reasoning, and speech interaction to support users in real-world tasks. CReLeRI is an explainable video analysis system that segments long videos, recognizes actions, and grounds them in the visual scene for human analysts. A comparative study of four action understanding paradigms — online detection, offline detection, sliding-window recognition, and MLLM-based video understanding — benchmarks their trade-offs in accuracy, efficiency, and data requirements on THUMOS’14 and ActivityNet. Finally, ELI-Mamba, an efficient long-range model for online action detection, matches state-of-the-art accuracy with far fewer parameters while running in real time, and will be released open-source.
Together, these contributions show that real-time action understanding depends on more than raw accuracy — it requires balancing localization, causality, efficiency, and system integration.
We are incredibly proud of Michael’s achievement and look forward to seeing where his research takes him next. Please join us in congratulating Dr. Michael Perez!