ECCV 2012. Workshops and Demonstrations

Abstract: In this study, we proposed and evaluated integrating local descriptors of images and 3D data in the context of indoor scene localization. We proposed normalization methods which are necessary and critical in information fusion. Since the RGB-D data contains complementary information with depth maps and RGB, their combination always introduces improvements in recognition. In our dataset, images belonging to same location is captured at the same time with similar illumination conditions. As a result, 2D descriptors achieved a considerably high recognition rate by themselves. We believe that 3D information will be more effective for datasets containing high illumination variations among images.