CDT Pizza 25/01

Talk #1

Speaker

Laura Sevilla-Lara

Title

Only time can tell: An overview of our work on temporal modeling for video understanding

Abstract

Deep learning has been incredibly useful in the last years to advance image understanding, and tasks like object segmentation are being solved with very high accuracy today. However, the video understanding problem which is the most natural for many artificial intelligence applications is lagging years behind. Some of the fundamental issues include: CNNs do not model well temporal information, computation is slower, methods rely on optical flow estimation, etc. In this talk I will describe our recent work to push the core problem in video understanding, which is temporal modeling. These involve integrating high-level recognition and low-level perception; introducing novel architectures to leverage temporal redundancy; and identifying the tasks that actually force intelligent systems to learn temporal information.