Predictive Visual Understanding

If you see a person start to sit down as someone else pulls away the chair, what could happen next? The ability to anticipate the events that may happen in the future, such as the person falling, is a key capability that will enable situated
machines to understand and interact with people in dynamic, real-world environments. However, while large annotated datasets fuel rapid advancements in visual recognition, machine understanding of events and dynamics remains challenging because the amount of knowledge required for understanding people in video is vast and potentially ambiguous. We aim to capitalize on large amounts of unlabeled video in order to develop artificial systems that learn to predict future and occluded behaviors of events and people.

Lab: Computer Vision Lab

Direct Supervisor: Carl Vondrick

Position Dates: 6/1/2019 - 9/1/2019

Hours per Week: 30

Paid Position: No

Credit: Yes

Number of positions: 1

Qualifications: Deep learning, machine learning, computer vision, big data

Eligibility: Sophomore, Junior (SEAS only)

Carl Vondrick, cv2428@columbia.edu