Learning for Object recognition in the RGB-D is a fundamental problem in robotics and computer vision community. One key problem with object detectors is that they work well on the data they are trained on but generalize poorly to data from other domains. In this context, a domain may refer to data of a certain type, from a certain source, or generated in a certain period in time. In practice, object detectors are often trained on a particular domain but in the application phase might be applied to another one. This degrades detector performance, a phenomenon that is commonly referred to as domain change problem. The domain change problem has recently been studied in the computer vision community, and approaches to tackle this problem by adapting pre-trained object detectors in images and videos to new domains have been proposed. However, little work has been done on understanding the effects of domain change on RGB-D data.
This work will investigate domain change problem in RGB-D data and will investigate approaches for domain adaptation – i.e. how can an RGB-D object detector trained on one domain be quickly adapted to the new domain by using a few labeled samples from that domain. Furthermore, a study of state-of-the-art RGB-D feature extractors / object classifiers will be performed to identify which combinations remain least affected by the domain change problem.
To apply, please email us with the following information:
Please email all files as PDF to: