Superpixel segmentation based gradient maps on RGB-D dataset

Superpixels aim to group homogenous pixels by a series of characteristics in an image. They decimate redundancy that may be utilized later by more computationally expensive algorithms. The most popular algorithms obtain superpixels based on an energy function on a graph. However, these graph-based methods have a high computational time consumption. This study presents a fast and high quality over-segmentation method by a watershed transform based on computing the dissimilarity of pixels among RGB(D) cues and gradient maps. Specifically, we first capture a gradient map based on an image to enhance and explain directional variations in the image scene. A distance function then measures the similarity among adjacent pixels, which is calculated according to RGB(D) values. A fast marker-controlled watershed (MCW) algorithm traverses the entire image based on the distance function. Finally, we acquire all watersheds consisting of superpixel contours. Experimental results compare state-of-the-art algorithms and highlight the effectiveness of the proposed method. As an application, the proposed superpixel algorithm can be used in applications aiming for real-time, like mobile robot saliency detection and segmentation.