Deep Relocalization with Metric Learning

Brigit Schroeder, Stanford Graduate Researcher in the Computational Vision and Geometry Lab (CVGL) talks about Deep Relocalization with Metric Learning.

Relocalization is one of the most challenging and key parts in real-time visual tracking, found in simultaneous localization and mapping (SLAM) systems. Tracking failure is a critical problem and a system′s ability to recover relies upon its ability to accurately recognize locations it has previously visited. An appearance-based relocalization system must be invariant to changes in viewpoint, illumination, and scale. In our work, we apply deep metric learning to the problem of appearance-invariant relocalization. We use a triplet convolutional neural network model to learn an embedding which projects images into a lower dimensional manifold. The network, trained with an indoor location-based dataset, is able to learn visual similarity and dissimilarity through the careful selection of triplets of image scenes. The compact embedding is able to successfully encode a large degree of appearance change for a specific location or area, making it an excellent representation of locality in a tracking system. The proposed approach uses a fast small deep learning model, making it more efficient than other SLAM-based relocalization methods.