Bayer feature map approximations through spatial pyramid convolution

摘要

Feature extraction is a key element of object detection and recognition. In Convolutional Neural Networks, feature maps are used to successively refine candidate features to ultimately determine classification results. When using raw bayer data, we show that efficient feature extraction can be achieved with just bayer data samples that are collected as green, red and blue samples arranged in a RGGB 2×2 pattern. Moreover, we show that spatial pyramid sampling using raw bayer sensor data can improve the efficiency for object classification in images.