Efficient Grapevine Structure Estimation in Vineyards Conditions
Developing computer vision systems for agricultural tasks that work in real-world conditions and in real time is challenging, especially if they need to be deployed on embedded devices, such as tablets or augmented reality glasses. In this paper, we present an efficient deep-learning approach for the estimation of grapevine structure in natural conditions with the aim of assisting vinemakers in some decision-making activities like grapevine pruning. Specifically, we propose a lightweight network for detecting nodes and branches in images which are then used to recover the tree structure. Our approach is validated on the publicly available 3D2Cut dataset. Compared to the ViNet method , we demonstrate computational performance while preserving the high accuracy of its predictions. Furthermore, we created a new dataset to train our workflow in real vineyard conditions without an artificial background. We demonstrate that we can obtain remarkable results in real and challenging conditions while being efficient.