Abstract: Most modern multi-instance pose estimation neural networks --either bottom-up or top-down variants-- are built around a highly specialized and constrained architecture. They rely on the detection of a fixed set of keypoints specific to the object in order to regress the pose in images. While efficient, those architectures are not very flexible and cannot be applied to objects with a less stable structure such as plants. In this paper we propose a neural network called SDNet which is suitable for the real-time detection of object poses with an unconstrained number of keypoints. To demonstrate this capability as well as its potential application for precision agriculture we evaluate it on a custom crop structure dataset and we compare its performance to the state-of-the-art neural network for real-time object detection Tiny YOLO v4 on two tasks where both of them can compete: (i) multi-instance crop detection and leaf counting --which can be applied to in-field phenotyping-- and (ii) stem and leaf keypoints detection and location --which can be used for real-time precision hoeing. We show that SDNet achieves excellent performance while still providing additional information via its unique structure detection ability.