CNN-CA: Convolutional Neural Network Combined with Active Contour for Image RGB-D Segmentation

Convolutional neural networks (CNNs) usually obtain the best results in image segmentation without a priori. However, they have two weaknesses: they do not take into account edge and local features. These weaknesses are even truer in the case of RGB-D (colour and depth) image segmentation because the complexity is higher. RGB-D segmentation CNNs have, at most, an average IoU below 60.0. To improve the result, the proposed new method is a combination of two existing methods: CNN and active contours. Active contours take into account the edges of objects by using customisable forces. These forces are specific to colour or depth images. CNN predicts the initial segmentation and active contours, using RGB-D forces, adjust the segmentation to the edges of the objects. The proposed method improves the CNN segmentation by 8.0 IoU on average. The proposed solution is effective in some cases. The exploration and analysis of the forces allow to find the essential forces to improve the segmentation of the objects. These forces have to be refined in order to have an optimal segmentation.