Taking Advantage of Depth Information for Semantic Segmentation in Field-Measured Vineyards

RGB-D cameras mounted on moving agricultural robotic platforms provide detailed information about both appearance and volume of plants. Those images can be analysed by means of deep segmentation models; however, such methods usually dismiss depth information. In this work, we aim to address this challenge by comparing four deep learning models for segmenting canopy and grape bunches in RGB and RGB-D images. In our experiments, RGB-D models achieved better results than their RGB counterparts, improving up to a 1.83% the mean segmentation accuracy. These findings highlight the potential of cost-effective RGB-based depth estimation techniques for accurate plant segmentation in agricultural settings, paving the way for wider adoption of RGB-D technology.