Accurate quantification of young fruit dimensions and precise localization of thinning points on fruit stalks represented core challenges for intelligent apple fruit thinning. To address this, a 3D recognition and sizing method integrating RGB D data with geometric constraints was developed for young apples and fruit stalks during the thinning period. The YOLO v8 architecture was enhanced through integration of a convolutional attention fusion module (CAFM) and adoption of the SIoU loss function, yielding the proposed YOLO YF detection model. Mask R-CNN was subsequently employed for high-precision instance segmentation of young fruits and fruit stalks. RGB-D image acquisition was performed by using an Intel RealSense D435i depth camera, followed by rigorous image alignment and 3D coordinate transformation. A spatial subordination discrimination algorithm based on geometric features was established to determine fruit stalk relationships. Experimental validation demonstrated that the YOLO-YF model achieved 86.80% precision, 78.70% recall, and 84.40% mAP50 in clustered young fruit clusters detection. Mask R-CNN attained 91.20% segmentation precision and 80.94% IoU, enabling reliable differentiation between fruit stalks and petioles. In 3D measurements, root mean square errors ( RMSE) of 1.43mm (radial) and 1.28 mm (axial) were obtained for young fruits. The positioning error at the optimal thinning point (fruit stalk midpoint) was constrained to 1.20 mm. This approach can provide a technical pathway characterized by quantitative sizing and millimeter-level positioning accuracy for intelligent thinning equipment. The methodology demonstrated extensibility to clustered fruit management applications.