Enhanced surface normal computation by exploiting RGB-D sensory information

Reliable surface normal computation is fundamental for a broad range of computer vision application areas, e.g. object segmentation, classification and recognition. Naturally, the surface normal is computed on the acquired depth data, whereby the normal quality is dependent on noise performance and resolution of the underlying image modality. The tendency of combining different imaging sensors into one device is increasing and leads to a new sampling density, which can be used to compensate or reduce the drawbacks of modalities. This paper presents a novel method for computing surface normals on RGB-D images by combining superpixel segmentation in color space with plane fitting in depth space. For evaluation we perform a qualitative comparison between our method and two standard methods for computing normal maps in diverse indoor scenes. Our results show an improvement in computing normal maps with clear and crisp-edges. Furthermore, our proposed method approximates valid normal information in areas where the depth sensor returned errors or depth inhomogeneities. These results emphasize our assumption that under normal light conditions edges in depth space are coherent to edges in color space.