Abstract:
This thesis primarily covers work on two different tasks in computer vision: (1) anomaly detection and (2) instance segmentation. Anomaly detection is an underexplored unsupervised problem that has existed in many fields. On the other hand, instance (and panoptic) segmentation is a supervised problem that can leverage the powerful data and key developments from the past few years.

We will first explore key challenges underlying anomaly detection and ways to overcome them. One of these key challenges lies in generating representations in which anomalies appear distinct, without training ahead of time on the task. To this aim, we’ll discuss two main approaches: (1) ‘dodging’ the issue by leveraging the discriminative formulation of anomaly detection and using other instances as context; and (2) tackling the issue head-on by working on object-centric pretext tasks that consider relationships between instances. Along with the second line of work, we will touch on results in instance and panoptic segmentation as well as a baseline for `relational anomalies’ based on scene graphs.

More Information

Thesis Committee Members:
Martial Hebert, Chair
J. Andrew Bagnell, Co-chair
David Held
Barnabas Poczos
Karteek Alahari, Inria Grenoble – Rhône-Alpe