SHOG - Spherical HOG Descriptors for Rotation Invariant 3D Object Detection

We present a method for densely computing local spherical histograms of oriented gradients (SHOG) in volumetric images. The descriptors are based on the continuous representation of the orientation histograms in the harmonic domain, which we compute very efficiently via spherical tensor products and the fast Fourier transformation. Building upon these local spherical histogram representations, we utilize the Harmonic Filter to create a generic rotation invariant object detection system that benefits from both the highly discriminative representation of local image patches in terms of histograms of oriented gradients and an adaptable trainable voting scheme that forms the filter. We exemplarily demonstrate the effectiveness of such dense spherical 3D descriptors in a detection task on biological 3D images. In a direct comparison to existing approaches, our new filter reveals superior performance.