Predicting the unseen: a new method for accurate yield estimation in viticulture/horticulture

"The project aims to revolutionise crop monitoring and yield estimation through the use of multi-camera 3D scanning to solve the significant problem of leaf-occlusion. Utilizing time series scans with modern differentiable rendering methods (Gaussian Splatting based) along with Functional Structural Plant Modelling (FSPM), to enable highly accurate yield prediction and understanding. We have built autonomous scanning vehicles with a multitude of sensors (GPS, LiDAR, IMU) synchronized with an array of cameras and flashes to reduce occlusion where single camera systems cannot, and use 3D through all parts of the pipeline to avoid ambiguity found in such monocular systems (such as double counting), and much more accurately understand what is visible. This is enabled through the use of very recent advances (and continued innovation) in differentiable rendering techniques, meaning we can very precisely reproduce views of a plant as well as extract the geometry to a much greater detail ever than before, and do so more robustly under outdoor lighting conditions, and model effects such as wind and camera defocus which were otherwise not feasible."