MVLoc: Multimodal Variational Geometry-Aware Learning for Visual Localization
share
Recent learning-based research has achieved impressive results in the field of single-shot camera relocalization. However, how best to fuse multiple modalities, for example, image and depth, and how to deal with degraded or missing input are less well studied. In particular, we note that previous approaches towards deep fusion do not perform significantly better than models employing a single modality. We conjecture that this is because of the naive approaches to feature space fusion through summation or concatenation which do not take into account the different strengths of each modality, specifically appearance for images and structure for depth. To address this, we propose an end-to-end framework to fuse different sensor inputs through a variational Product-of-Experts (PoE) joint encoder followed by attention-based fusion. Unlike prior work which draws a single sample from the joint encoder, we show how accuracy can be increased through importance sampling and reparameterization of the latent space. Our model is extensively evaluated on RGB-D datasets, outperforming existing baselines by a large margin.
READ FULL TEXT