Quaternionic keypoint description for multispectral imaging

In this work, we propose a new keypoint descriptor that is suitable for multispectral inputs comprising up to 4 channels. Color images with or without an additional infrared or depth channel are some of the use-cases that can be handled by the proposed descriptor. Standard keypoint descriptors employ single-channel input gradients, thereby discarding potentially useful content. The proposed descriptor is based on a quaternionic representation of the input image thereby treating each pixel multispectral value holistically. Coupled with a suitable multispectral quaternionic detector, we show that the proposed detector leads to superior experimental results on a keypoint matching scenario.