Spherical CNNs (2018)
a year ago
- #Computer Vision
- #Spherical CNNs
- #Machine Learning
- Spherical CNNs are introduced to analyze spherical images, addressing limitations of traditional CNNs with planar images.
- Applications include omnidirectional vision for drones, robots, autonomous cars, molecular regression, and climate modeling.
- A naive approach of applying CNNs to planar projections of spherical signals fails due to space-varying distortions.
- The paper proposes a spherical cross-correlation definition that is expressive and rotation-equivariant.
- Efficient computation is enabled via a generalized Fast Fourier Transform (FFT) algorithm.
- Demonstrated effectiveness in 3D model recognition and atomization energy regression.