LeafTrackNet: A Deep Learning Framework for Robust Leaf Tracking in Top-Down Plant Phenotyping

Leaf-level phenotyping can reveal early signals of plant growth and stress, making it a key step toward understanding crop development. However, tracking individual leaves over time is still challenging, especially for structurally complex crops such as canola. This difficulty stems both from the scarcity of realistic, publicly available benchmark datasets and from the limitations of current methods: existing plant-specific tracking methods often rely on Intersection-over-Union (IoU) thresholds to associate leaves between frames, which can break down when leaves overlap, grow, or change shape. Generic multi-object tracking (MOT) methods, on the other hand, are designed for approximately rigid objects like cars or pedestrians and struggle with the continuous deformation and complex motion patterns of leaves. Therefore, the contribution of our work is two-folded - First, we introduce $\textbf{CanolaTrack}$, a high-resolution dataset of 5,704 top-down RGB images with 31,840 annotated leaf instances spanning the early growth stages of 184 canola plants. Second, we propose $\textbf{LeafTrackNet}$, an efficient lightweight framework for long-term leaf tracking. It combines a YOLOv10 detector with a MobileNetV3 embedding head and links identities via cosine similarity and Hungarian assignment, without geometric motion priors. On CanolaTrack, LeafTrackNet outperforms both plant-specific tracking methods and state-of-the-art MOT baselines, improving HOTA by 9.73%. Our work provides a realistic benchmark dataset and a simple, effective framework for long-term leaf tracking, contributing to AI-driven plant phenotyping. Code and dataset are available at https://github.com/shl-shawn/LeafTrackNet.