Tomato Fruit 3D Reconstruction Based on Improved YOLOv8 and MiDaS Depth Estimation

In order to address the challenges of inaccurate crop phenotypic parameters and unclear crop growth status in crop yield prediction, this study focuses on tomato fruits in greenhouse production environments. It delves into the three-dimensional reconstruction of multi-view images of tomato fruits. This paper introduces an enhanced YOLOv8 tomato fruit recognition model to address the challenge of overlapping fruits, leaves, and stems in complex background environments. By incorporating the CBAM attention mechanism into the model backbone, the recognition mAP@0.5 index improves to 99.3%, a 1.9% increase from the original model. Three-dimensional reconstruction technology is currently a research focus in the field of crop yield prediction. The three-dimensional structure of a crop is a reflection of its growth status. By establishing an accurate three-dimensional crop model, researchers can analyze crop organs to obtain phenotypic parameters. The MiDaS algorithm estimates depth, and the Wonder3D algorithm achieves three-dimensional reconstruction of single-view images. These methods offer an effective approach for reconstructing tomato fruits in three dimensions from multiple viewpoints, with potential applications in tomato yield prediction and the development of harvesting robots.