Abstract: The leaf count of tobacco plants is one of the important phenotypic parameters for tobacco leaf yield estimation. To address the challenges of traditional manual tobacco leaf counting, a field tobacco leaf counting method integrating three-dimensional point clouds and improved PointNet + + was proposed. This method employs UAV oblique photography to acquire field tobacco plant images and generate three-dimensional point clouds. An improved PoinNnet ++ algorithm is then utilized to perform leaf point cloud segmentation. The proposed algorithm replaces the MLP with KAN to enhance learning capacity and minimize training loss. A DGSTD attention mechanism is proposed, which integrates DGST network and DBB multi-branch block to enhance accuracy. Additionally, Varifocal loss is incorporated to address the class imbalance in point cloud distribution across categories. Finally, the MeanShift clustering algorithm is employed to cluster the leaf point clouds, from which the leaf count is derived. The results show that the accuracy of point cloud segmentation is 92.55%, and the mean Intersection over Union (mIoU) is 76.33%, representing improvements of 2.06 and 2.81 percentage points over the original model, respectively. The proposed method achieves a leaf counting precision of 94.35%, successfully implementing leaf counting of field tobacco plants in three-dimensional space.