Abstract:
We investigate a geometrically motivated modification to semantic segmentation. In particular, we reformulate typical planar CNN as a projected spherical CNN where image distortions are reduced, and thus generalisation increased. Since prior formulations of spherical CNNs require computation on full spheres, fair comparison between planar and spherical methods have not been previously presented. In this work, we first extend spherical deep learning to support high-resolution images by exploiting the reduced field of view of classical images. Then, we employ our spherical representation to reduce distortion effects of standard deep learning systems. On typical benchmarks, we apply our spherical representation and consistently outperform the classical representation of multiple existing architectures. Additionally, we introduce direct spherical pretraining from planar datasets to further improve results. Finally, we compare our method on non-planar datasets, where we improve accuracy, and outperform running time of spherical state of the art for non-complete input spheres.
