OWOD: Towards Open World Object Detection

April 2021

tl;dr: Propose a new task of open world object detection. Detect unknown objects, while able to continuous benefit from new labels on these unknown objects.

Overall impression

All existing object detectors have one strong assumption that all the classes are to be detected would be available at training phase. Although there are open set classification problem, the object detection is not studied. Object detectors are trained to detect unknown objects as background.

The open world object detector is required to detect unknown classes as unknown, and is able to detect these classes when the unknown instances are forwarded to a human annotator to label, without being trained from scratch.

Key ideas

Technical details