Fruit Size Estimation by Integrating Depth Data Calculated Using Monocular Images

Abstract

Precise estimation of fruit size is important for yield estimation and autonomous harvesting. We introduce a technique to estimate fruit sizes from monocular RGB images, bypassing the limitation of object detection models such as Faster R-CNN, which fail to classify size. Our technique classifies fruit as small, medium, and large from RGB input and boosts accuracy with depth information from MiDaS, a monocular depth estimation model. We have tried ensemble learning, clustering, and artificial neural networks (ANN) to determine the effect of depth. Experiments using a self-assembled mango dataset show that a Random Forest classifier that includes depth features has an accuracy of 67.47\%, outperforming traditional methods. The results show that depth information consistently enhances classification performance in all models studied. The research contributes to the development of precision agriculture by allowing automatic fruit size classification without the need for specialized depth-sensing hardware.

Keywords: Fruit size estimation, Monocular depth, Object Detection, precision agriculture