CanolaTrack: Canola Phenotyping Benchmark

• CanolaTrack is a domain-specific benchmark for canola phenotyping that maintains persistent identity tracking of leaves and branches across realistic, non-ideal imaging conditions.
• It integrates multiple tracking modalities, including top-down leaf tracking, branch identity preservation, and 3D canopy reconstruction to support longitudinal crop analysis.
• Benchmark outcomes show significant improvements over standard MOT methods, with metrics such as HOTA, MOTA, and IDF1 demonstrating its robust performance.

Taken together, papers suggest that CanolaTrack has a dual status in current canola phenotyping literature. In its most specific sense, it denotes a domain-specific benchmark for leaf-level multi-object tracking in top-down RGB imagery of Brassica napus. In a broader, task-oriented sense, the name is also used for canola-centric tracking pipelines spanning branch identity preservation, row-wise plant MOTS for precision spraying, growth-stage estimation, plot-scale canopy reconstruction, clubroot diagnosis, and multitemporal crop mapping. Across these usages, the common technical problem is longitudinal estimation under non-rigid growth, occlusion, temporal sparsity, and biologically driven appearance change rather than rigid-body motion alone (Liu et al., 15 Dec 2025, Chen et al., 17 Dec 2025, Hu et al., 2023, Ngorima et al., 10 Jun 2026, Guo et al., 23 Jun 2025, Zhu et al., 19 Jan 2026, Jacques et al., 2023).

1. Conceptual scope and phenotyping motivation

Canola tracking is motivated by the agronomic value of persistent, organ-resolved measurements over time. Leaf-level phenotyping captures emergence, growth, morphological change, occlusion, and senescence; branch-level tracking supports longitudinal traits such as elongation rate, angle trajectories, and flowering duration; plot-scale reconstruction enables canopy architectural indicators and yield-related proxies. These targets are biologically meaningful because developmental and stress responses are expressed as temporally evolving structures rather than isolated snapshots (Liu et al., 15 Dec 2025, Chen et al., 17 Dec 2025, Guo et al., 23 Jun 2025).

The central technical difficulty is that canola violates the assumptions underlying many generic tracking systems. In top-down rosette imagery, leaves exhibit large intra-plant similarity, non-rigid deformation due to growth, frequent self-occlusions, and occasional global pot rotations that break geometric continuity. In side-view canopy imagery, branches become entangled, stage-dependent ambiguity increases, and identities fragment as the canopy densifies. In field-scale sensing, cloud cover, temporal gaps, and mixed pixels complicate multitemporal classification. This suggests that CanolaTrack-style systems are defined less by a single sensor or architecture than by a shared requirement: maintaining biological identity or state through realistic, non-ideal imaging regimes (Liu et al., 15 Dec 2025, Chen et al., 17 Dec 2025, Jacques et al., 2023).

2. The leaf-level CanolaTrack benchmark

The most concrete and fully specified instantiation of CanolaTrack is the benchmark introduced with LeafTrackNet. It comprises 5,704 top-down RGB images at 1200×1200 resolution, collected daily over 31 consecutive days for 184 individual canola plants, yielding 184 sequences with 31 frames each and $\Delta t = 24$ hours. The dataset contains 31,840 annotated leaf instances in total, each linked to a persistent identity maintained across the sequence. The reported dynamics include leaf birth, non-uniform expansion, overlap, occlusion, disappearance through death or senescence, re-appearance after occlusion, and occasional whole-pot rotation (Liu et al., 15 Dec 2025).

Component	Value	Notes
Images	5,704	Top-down RGB, 1200×1200
Plants / sequences	184	31 daily frames per plant
Leaf instances	31,840	Persistent identity annotations
Train split	147 plants	4,557 images; 25,485 leaves
Test split	37 plants	1,147 images; 6,355 leaves

The benchmark is explicitly positioned against prior plant datasets such as LeTra, KOMATSUNA, and MSU-PID. It is described as a substantially larger, high-resolution, RGB-only corpus captured under realistic, temporally sparse conditions, and as the largest dataset for leaf tracking in agriculture crops (Liu et al., 15 Dec 2025). The intended tasks are leaf detection in top-down RGB imagery and multi-object leaf tracking with persistent identities across days. Annotation is bounding-box based using top-left, width, height; masks and keypoints are not provided. Identities are maintained across frames within a plant sequence, but annotation tools, inter-annotator agreement, and quality-control procedures are not reported; annotation was performed by one of the authors (S.L.) (Liu et al., 15 Dec 2025).

Benchmarking follows standard MOT practice via TrackEval, with HOTA, DetA, AssA, MOTA, and IDF1. The benchmark defines

$$\mathrm{HOTA} = \sqrt{\mathrm{DetA} \times \mathrm{AssA}}$$

and

$$\mathrm{DetA} = \frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FP} + \mathrm{FN}}.$$

This metric set is significant because the dataset is designed not merely to reward accurate per-frame detection but to measure long-horizon identity preservation under daily sampling and strong non-rigid change (Liu et al., 15 Dec 2025).

3. LeafTrackNet: architecture, association mechanism, and reported performance

LeafTrackNet is the benchmark’s reference framework. It decouples detection from identity association and replaces explicit motion modeling with memory-based embedding association. The detector is a YOLOv10 model fine-tuned on CanolaTrack; at inference, detections with confidence $< 0.5$ are filtered. Each detection is cropped and resized via

$$x_{k,i}^t = \psi(I_k^t, b_{k,i}^t) \in \mathbb{R}^{3 \times W \times H},$$

with $W = H = 224$ in all experiments. The embedding network uses a MobileNetV3 backbone $N_\theta$, pretrained on ImageNet and truncated before classification, followed by a linear projection $F_\phi$:

$E(x) = F_\phi(N_\theta(x)) \in \mathbb{R}^D.$

The paper does not specify $D$ (Liu et al., 15 Dec 2025).

Embedding learning is organized by leaf-level triplets. For anchor $$\mathrm{HOTA} = \sqrt{\mathrm{DetA} \times \mathrm{AssA}}$$0, the positive is $$\mathrm{HOTA} = \sqrt{\mathrm{DetA} \times \mathrm{AssA}}$$1 with $$\mathrm{HOTA} = \sqrt{\mathrm{DetA} \times \mathrm{AssA}}$$2, and the negative is $$\mathrm{HOTA} = \sqrt{\mathrm{DetA} \times \mathrm{AssA}}$$3 with $$\mathrm{HOTA} = \sqrt{\mathrm{DetA} \times \mathrm{AssA}}$$4, taken from a different leaf of the same plant. The triplet margin loss uses $$\mathrm{HOTA} = \sqrt{\mathrm{DetA} \times \mathrm{AssA}}$$5:

$$\mathrm{HOTA} = \sqrt{\mathrm{DetA} \times \mathrm{AssA}}$$6

Training uses Adam with lr $$\mathrm{HOTA} = \sqrt{\mathrm{DetA} \times \mathrm{AssA}}$$7, weight decay $$\mathrm{HOTA} = \sqrt{\mathrm{DetA} \times \mathrm{AssA}}$$8, batch size 48, up to 80 epochs with early stopping, on four NVIDIA Tesla V100S GPUs. Data augmentation is not described (Liu et al., 15 Dec 2025).

At inference, identity is preserved purely through embedding similarity and temporal smoothing. At $$\mathrm{HOTA} = \sqrt{\mathrm{DetA} \times \mathrm{AssA}}$$9, each detection initializes a prototype in a memory bank. At later times, LeafTrackNet computes cosine similarities

$$\mathrm{DetA} = \frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FP} + \mathrm{FN}}.$$0

forms $$\mathrm{DetA} = \frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FP} + \mathrm{FN}}.$$1, and solves one-to-one assignment with the Hungarian algorithm. Matches are accepted when $$\mathrm{DetA} = \frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FP} + \mathrm{FN}}.$$2 with default $$\mathrm{DetA} = \frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FP} + \mathrm{FN}}.$$3. Prototypes are updated by exponential moving average,

$$\mathrm{DetA} = \frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FP} + \mathrm{FN}}.$$4

with $$\mathrm{DetA} = \frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FP} + \mathrm{FN}}.$$5; unmatched tracks age and are removed when $$\mathrm{DetA} = \frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FP} + \mathrm{FN}}.$$6 with default $$\mathrm{DetA} = \frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FP} + \mathrm{FN}}.$$7. No motion model or IoU gating is used (Liu et al., 15 Dec 2025).

The reported benchmark outcome is unambiguous. LeafTrackNet surpasses both plant-specific trackers and state-of-the-art MOT baselines on CanolaTrack, achieving HOTA 88.03, compared with 78.30 for the best baseline MOTRv2; DetA 92.25 versus 91.94 for ByteTrack; AssA 84.07 versus 79.36 for MOTRv2; MOTA 93.64 versus 82.09 for LeTra; and IDF1 92.90 versus 83.78 for MOTRv2. The abstract summarizes this as a 9% HOTA improvement. The general-purpose baselines BoT-SORT and ByteTrack show high DetA near 92 but weak AssA near 12, indicating that geometric association and rigid-motion models fail under occlusion, growth-induced deformation, and rotation (Liu et al., 15 Dec 2025).

Ablations reinforce the design logic. Replacing MobileNetV3 with ResNet-18/34/50/101 or ViT-B/16 does not improve identity preservation; MobileNetV3 gives the best or near-best HOTA/AssA/IDF1 while remaining the most efficient at approximately 2.97M parameters and 0.23G MACs. HOTA remains stable for $$\mathrm{DetA} = \frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FP} + \mathrm{FN}}.$$8, and moderate EMA smoothing around $$\mathrm{DetA} = \frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FP} + \mathrm{FN}}.$$9 works best. Qualitative analyses show robustness to dense occlusions of small leaves and to large, non-physical motions such as 90° pot rotations between days. The reported failure modes of IoU/Kalman trackers and orientation-sensitive ReID systems are therefore not incidental but structural consequences of the imaging regime (Liu et al., 15 Dec 2025).

4. Branch identity and plant-instance tracking variants

A second usage treats CanolaTrack as a branch identity tracking pipeline, operationalized by ST-DETrack. Here the task is to assign each floral bud $< 0.5$0 to its originating branch point $< 0.5$1 and preserve that mapping from budding through flowering and pod fill. The method uses a shared ViT-B/16 encoder with two specialized decoders: a spatial decoder that leverages position and angle, and a temporal decoder that exploits motion consistency. Their outputs are fused by an adaptive gate $< 0.5$2, with a negative gravitropism prior penalizing implausible downward matches (Chen et al., 17 Dec 2025).

The branch-tracking dataset comprises 360 plants, three side views (0°, 45°, 90°), 4,877 frames, 22,911 annotated branch points, and 38,187 annotated buds. Images are standardized by cropping background and resizing to 640×640 and then 224×224; a binary plant mask generated via Li’s minimum cross-entropy thresholding is concatenated with RGB. Identities are maintained using order-based semantic indices, and temporal models consume an 8-channel tensor formed by previous and current RGB+mask frames. Reported performance is BMA = 93.6%, MOTA = 0.980, IDF1 = 0.993, FP = 0, and FN = 5. Relative to the spatial-only baseline (BMA 64.7%), ST-DETrack improves by 28.9 percentage points; relative to the temporal-only baseline (BMA 90.3%), it improves by 3.3 points (Chen et al., 17 Dec 2025).

The score fusion is explicitly stage-dependent:

$< 0.5$3

with $< 0.5$4 and $< 0.5$5. This formalizes the observation that early sparse stages are primarily resolvable by geometry, whereas late dense stages require temporal continuity. The paper also reports LIoU = 0.9617 and BTC = 0, indicating robust longitudinal skeleton fidelity (Chen et al., 17 Dec 2025).

A different branch of the literature addresses plant-instance tracking for robotic precision spraying. The MOTS method in "Segmentation and Tracking of Vegetable Plants by Exploiting Vegetable Shape Feature for Precision Spray of Agricultural Robots" was tested on lettuce, but the authors state that it can be applied to similar vegetables such as broccoli and canola. Its association logic combines contour Fourier descriptors, ellipse blob features, a Kalman filter with state

$< 0.5$6

and Hungarian assignment with cost

$< 0.5$7

where $< 0.5$8 is cosine distance between shape features and $< 0.5$9 is a GIoU-based position cost. On the LettuceMOTS test sequences, the method reports HOTA 71.99, 72.08, and 70.10, with end-to-end runtime around 34 ms per frame or approximately 29 FPS on NVIDIA RTX 2080Ti + Intel i7-10700K (Hu et al., 2023).

These branch and plant-instance variants show that CanolaTrack is not reducible to one canonical architecture. In side-view branch tracking, explicit geometric priors and stage-adaptive fusion are central. In row-wise MOTS, contour shape and local motion remain important. In top-down leaf tracking, by contrast, memory-based appearance embeddings outperform rigid-motion association. This suggests that the relevant inductive bias is view- and organ-specific rather than universal (Chen et al., 17 Dec 2025, Hu et al., 2023, Liu et al., 15 Dec 2025).

5. Developmental, pathological, and 3D phenotyping extensions

One developmental extension treats CanolaTrack as growth-stage estimation. In "Feature extraction for plant growth estimation", canola and radish are classified into background plus four growth stages using indoor, top-mounted imagery from the bccr-segset dataset. For canola, the total count is 5,378 images, with 4,831 for training and 547 for testing; imaging resolution is 228×228 pixels. Two feature pipelines are compared: a Gabor filter bank + morphological operations pipeline and a CNN feature extraction pipeline using VGG-16, VGG-19, or ResNet-50. The best reported system, VGG-19 features + RBF SVM, achieves 98.4% accuracy for canola and processes an image in 0.08 seconds. Misclassifications are concentrated between stages 1 and 2, while background and stage 4 are clearly separable (Ngorima et al., 10 Jun 2026). A plausible implication is that growth-stage labels can be appended to identity trajectories in a longitudinal CanolaTrack system, but the paper does not report temporal smoothing or sequence-level tracking.

A structurally different extension is plot-scale 3D canopy reconstruction via RP-PCN. This work addresses rapeseed population point cloud completion from seedling, bolting, flowering, and silique stages using multi-view imaging, virtual-real integration (VRI), automated occlusion annotation via ray tracing, and a GAN-based completion network with a Multi-Resolution Dynamic Graph encoder (MRDG), Dynamic Graph Convolutional Feature Extractor (DGCFE), and Point Pyramid Decoder (PPD). Field plots comprise 300 accessions, 3 replicates, or 900 plots, with oblique imaging from one circular UAV flight at −60° and 36 images per plot at 8000×6000 pixels. RP-PCN reports Chamfer Distance values of 3.35 cm, 3.46 cm, 4.32 cm, and 4.51 cm at seedling, bolting, flowering, and silique, respectively. Ablations show that MRDG reduces CD by 10% and DGCFE by 23%. At silique stage, the silique efficiency index (SEI) derived from completed point clouds improves yield prediction, increasing $$x_{k,i}^t = \psi(I_k^t, b_{k,i}^t) \in \mathbb{R}^{3 \times W \times H},$$0 from 0.81 to 0.90 and improving overall prediction accuracy by 11.2% (Guo et al., 23 Jun 2025).

Pathological monitoring introduces yet another modality. In "Non-Invasive Diagnosis for Clubroot Using Terahertz Time-Domain Spectroscopy and Physics-Constrained Neural Networks", THz-TDS is applied to canola leaves and roots as a non-contact, non-destructive, preparation-free diagnosis tool. The system uses a transmission-mode THz-TDS spectrometer with bandwidth exceeding 6 THz, dynamic range >100 dB and up to 110 dB, spectral resolution <0.6 GHz, and lateral scan step 0.5 mm. Refractive index and absorption coefficient are extracted as

$$x_{k,i}^t = \psi(I_k^t, b_{k,i}^t) \in \mathbb{R}^{3 \times W \times H},$$1

and

$$x_{k,i}^t = \psi(I_k^t, b_{k,i}^t) \in \mathbb{R}^{3 \times W \times H},$$2

Infected canola leaves exhibit a clearer low-frequency refractive-index blue shift than Arabidopsis leaves, while infected roots show a strong absorption band around 0.4 THz and consistently lower $$x_{k,i}^t = \psi(I_k^t, b_{k,i}^t) \in \mathbb{R}^{3 \times W \times H},$$3 across the measured band. A physics-constrained neural network with encoder channels 1→16→32→64→128, latent dimension 32, batch size 32, and 50 epochs is used for self-supervised feature extraction. The paper reports improved latent-feature contrast relative to PCA and a conventional autoencoder, but does not report accuracy, sensitivity, specificity, or ROC/AUC (Zhu et al., 19 Jan 2026).

6. Field-scale monitoring, limitations, and future directions

At the largest spatial scale, CanolaTrack intersects with multitemporal remote sensing. The Canadian Cropland Dataset contains 78,536 geo-referenced Sentinel-2 instances from 2017–2020 and June–October, with canola as one of 10 classes and 11,366 canola-labeled patches. Each instance covers 640×640 m at 10 m/pixel and includes 12 spectral bands, RGB, and five vegetation indices: NDVI, NDVI45, GNDVI, PSRI, and OSAVI. Sequence models outperform static image models: on the 2019 subset, LRCN-64 (RGB) reaches Accuracy 0.774, Precision 0.610, Recall 0.637, F1 0.617, and 3D-CNN (RGB) reaches Accuracy 0.773, while the best static model, ResNet-50 (RGB), reports Accuracy 0.667 (Jacques et al., 2023). This indicates that temporal context remains decisive even when tracking is reframed as crop-class persistence across months rather than organ identity across days.

Across the literature, several limitations are explicit. The leaf-level CanolaTrack benchmark is confined to early growth stages, top-down RGB imaging, and daily temporal sparsity; it does not include segmentation masks, depth, or physiological modalities, and it does not report cross-species generalization (Liu et al., 15 Dec 2025). ST-DETrack addresses only the reported Brassica napus side-view setting, and the paper does not provide a public code repository link or licensing terms (Chen et al., 17 Dec 2025). The THz-TDS study emphasizes feature extraction and contrast enhancement rather than supervised diagnosis, so operational detection thresholds and classification metrics remain unspecified (Zhu et al., 19 Jan 2026). The Canadian cropland benchmark does not provide per-class canola precision, recall, or confusion matrices, and label quality is bounded by the AAFC Canadian Annual Crop Inventory and the 10 m/30 m resolution mismatch (Jacques et al., 2023).

The forward trajectory is correspondingly multi-scale. The leaf-tracking benchmark identifies cross-species transfer, field-level deployment, real-time decision support, and extensions with denser temporal sampling, segmentation masks, and complementary sensing such as depth or fluorescence as natural next steps (Liu et al., 15 Dec 2025). Branch tracking points toward 3D/multi-view fusion, end-to-end joint detection + tracking, and plant topology priors (Chen et al., 17 Dec 2025). Plot-scale reconstruction highlights sensor fusion, uncertainty estimation, and plot-scale mapping with SLAM or odometry (Guo et al., 23 Jun 2025). The most defensible synthesis is therefore that CanolaTrack names an emerging canola phenotyping agenda: persistent, biologically faithful estimation across organs, stages, sensing modalities, and spatial scales, with identity preservation as its recurring methodological core.