Horticultural Temporal Fruit Tracking via 3D Instance Segmentation and Re-Identification

Robotic fruit monitoring is crucial for automating agricultural systems, offering precise, high-throughput assessments that address the limitations of manual methods. Monitoring is challenging due to variations in fruit size, shape, orientation, and occlusion. Most methods rely on 2D images, that lack 3D structure and depth. While 3D colored point clouds provide this information, they introduce challenges like sparsity and irregularity. We propose a novel method for temporal fruit monitoring using point clouds collected over time. Experiments on a strawberry dataset show that this method outperforms baselines for temporal fruit monitoring in real-world scenarios.