Slicing aided large scale tomato fruit detection and counting in 360-degree video data from a greenhouse

Abstract

This paper proposes an automated tomato fruit detection and counting process without a need for any human intervention. First of all, wide images of whole tomato plant rows were extracted from a 360-degree video taken in a greenhouse. These images were utilized to create a new object detection dataset. The original tomato detection methodology uses a deep CNN model with slicing-aided inference. The process encompasses two stages: first, the images are cut into patches for object detection, and consequently, the predictions are stitched back together. The paper also presents an extensive study of post-processing parameters needed to stitch object detections correctly, especially on the patch's borders. Final results reach 83.09% F1 score value on a test set, proving the suitability of the proposed methodology for robotic farming.