The US tree fruit industry is undergoing structural changes driven by shifting market conditions, rising production costs, and growing demand for high-quality fresh fruit, leading to increased adoption of modern high-density orchard systems. These plantings use trellis training to form compact, planar canopies that improve light interception, resource efficiency, fruit quality, and worker access. However, the plantings also demand intensive canopy management, which remains highly labor-dependent at a time when the agricultural workforce is limited and labor costs are increasing. Although the uniform structure of high-density canopies provides a more automation-friendly environment, a successful agricultural robot ultimately hinges on reliable, vision-based perception capable of extracting actionable information from complex orchard scenes. While perception systems for robotic pruning have progressed from handcrafted image-processing workflows to modern deep learning-based approaches, current methods still fall short of providing the fine-grained structural understanding required for making consistent pruning decisions. Latency also remains a key barrier to practical deployment. These limitations highlight the need for pipelines that are both fast and robust under real orchard conditions. This study introduces a deep learning-based perception pipeline that predicts pruning points in 3D from color image–depth map pairs captured by a depth camera from trees trained to the mini V-trellis orchard system. To reduce scene variability, experimental units were defined using the repetitive branch patterns of the orchard system. The camera was positioned consistently over these units, and branch structures were simplified to remove visual clutter and branching forks. Models from three instance segmentation architectures (Mask R-CNN, YOLO11, and RF-DETR) were trained to detect branches and compared. RF-DETR Seg achieved the highest mAP@50:95 (0.670) with a latency of 38.0 ms on the test hardware. A pruning rule for the mini V-trellis system that prescribes the removal of long shoots was then applied to the segmented branches to determine pruning points. The pipeline demonstrated strong end-to-end performance: pruning points were predicted from the color image–depth map pairs with high precision (94.5%) and recall (76.8%), for an F1 score of 84.7%, and with an overall latency of 51.8 ms. Failure mode analysis showed that 71% of pruning-point prediction errors stemmed from missed shoot instances in the instance segmentation module, while a skeletonization module effectively mitigated missing depth values. Overall, the workflow delivered fast and reliable pruning-point predictions, demonstrating its promise to provide a robust vision component for robotic pruning.