Alexander C Berg, Associate Professor at the University of North Carolina at Chapel Hill and CTO of Shopagon, presents the “Recognizing Novel Objects in Novel Surroundings with Single-shot Detectors” tutorial at the May 2018 Embedded Vision Summit.
Berg’s group’s 2016 work on single-shot object detection (SSD) reduced the computation cost for accurate detection of object categories to be in the same range as image classification, enabling deployment of general object detection at scale. Subsequent extensions add segmentation and improve accuracy, but still require many training examples in real-world contexts for each object category.
In certain applications, it may be desirable to detect new objects or categories for which many training examples are not readily available. Berg’s presentation considers two approaches to address this challenge. The first takes a small number of examples of objects not in context and composes them into scenes in order to construct training examples. The other approach learns to detect objects that are similar to a small number of target images provided during detection and does not requiring retraining the network for new targets.