In-field apple size estimation using photogrammetry-derived 3D point clouds: Comparison of 4 different methods considering fruit occlusions

In-field fruit monitoring at different growth stages provides important information for farmers. Recent advances have focused on the detection and location of fruits, although development accurate size estimation systems is still a challenge that requires further attention. This work proposes novel methodology automatic in-field apple which based four main steps: 1) detection; 2) point cloud generation using structure-from-motion (SfM) multi-view stereo (MVS); 3) estimation; 4) visibility estimation. Four techniques were evaluated in step. The first consisted obtaining diameter by measuring two most distant points an (largest segment technique). second third fitting sphere to least squares (LS) M−estimator sample consensus (MSAC) algorithms, respectively. Finally, template matching (TM) was applied 3D model points. best results obtained with LS, MSAC TM techniques, showed mean absolute errors 4.5 mm, 3.7 mm 4.2 coefficients determination (R2) 0.88, 0.91 Besides size, proposed method also estimated percentage apples detected. step R2 0.92 respect ground truth visibility. allowed identification discrimination measurements highly occluded apples. disadvantage high processing time required (in this 2760 s modelling 6 trees), limits its direct application large agricultural areas. code dataset been made publicly available visualization accessible http://www.grap.udl.cat/en/publications/apple_size_estimation_SfM.