Abstract:
Although computer vision has seen significant advances in static image analysis, the relatively slow advances in video tasks such as action detection suggest we’re struggling to build effective temporal models. In this talk, I will present a few main ideas that drive contemporary approaches, such as “two-stream networks” and “3D” convolutional networks. I’ll also present our work on a lightweight recurrent model we call “Predictive Corrective Networks”, and discuss future directions towards building more accurate models for action detection in videos.

Research Qualifier Committee:
Deva Ramanan
Abhinav Gupta
Katerina Fragkiadaki
Gunnar Atli Sigurdsson