Abstract

Over the last several years it has been shown that image-based object detectors are sensitive to the training data andoften fail to generalize to examples that fall outside the original training sample domain (e.g., videos). A number ofdomain adaptation (DA) techniques have been proposed toaddress this problem. DA approaches are designed to adapta fixed complexity model to the new (e.g., video) domain.We posit that unlabeled data should not only allow adap-tation, but also improve (or at least maintain) performanceon the original and other domains by dynamically adjust-ing model complexity and parameters. We call this notiondomain expansion. To this end, we develop a new scalableand accurate incremental object detection algorithm, basedon several extensions of large-margin embedding (LME).Our detection model consists of an embedding space andmultiple class prototypes in that embedding space, that rep-resent object classes; distance to those prototypes allowsus to reason about multi-class detection. By incrementallydetecting object instances in video and adding confident de-tections into the model, we are able to dynamically adjustthe complexity of the detector over time by instantiating new prototypes to span all domains the model has seen. We test performance of our approach by expanding an object detector trained on ImageNet to detect objects in egocentric videos of Activity Daily Living (ADL) dataset and challenging videos from YouTube Objects (YTO) dataset.