PhenoLeaf‑TS

A Time-Series Benchmark for Leaf Instance Segmentation, Tracking, and Growth Stage Classification

European Conference on Computer Vision (ECCV) 2026

Rijad Sarić^1,* Basim Azam² Sarmad Khan³ Edhem Čustović⁴

¹ ARC Research Hub for Protected Cropping (PC Hub), La Trobe University, Melbourne, Australia

² School of Computing and Information Systems, The University of Melbourne, Melbourne, Australia

³ Department of Computer Science, Technological University Dublin, Ireland

⁴ Photon Systems Instruments (PSI), Drásov, Czech Republic

^* Corresponding author

View Paper Code Dataset BibTeX

Overview

Image-based high-throughput plant phenotyping benefits greatly from an instance-level understanding of how individual leaves grow over experimental time. Yet existing datasets lack the temporal depth and annotation consistency needed to jointly benchmark segmentation, tracking, and growth stage classification.

We introduce PhenoLeaf-TS, a time-series dataset of 17,082 top-down RGB images spanning 21 Arabidopsis thaliana genotypes, totalling 318 plant replicates, each annotated with colour-coded leaf instance masks that maintain a consistent identity throughout the growth sequence. We define three benchmark computer-vision tasks with standardised protocols and evaluate 21 distinct models: 9 instance-segmentation architectures, 6 multi-object trackers, and 6 classification architectures.

Highlights

17,082 RGB images · 21 Arabidopsis genotypes · 318 plant replicates · ~530×525 px
Temporally consistent, colour-coded leaf instance masks (fixed 31-colour palette)
Three tasks in one dataset: instance segmentation, leaf tracking, growth stage classification
21 baseline models benchmarked under standardised protocols
First dataset to combine large-scale time-series imagery, per-leaf instance labels, tracking, and growth stages
Demonstrated value as a pre-training source: up to +51.4 mAP fine-tuning gain on external datasets

Overview of PhenoLeaf-TS: representative samples across genotypes and a temporal sequence of a single plant with persistent leaf colours. — **Figure 1.** Overview of PhenoLeaf-TS. (a) Representative samples across seven genotypes: raw RGB (first row), colour-coded leaf instance labels (second row), and overlays (third row). (b) Temporal sequence of a single plant (Day 1–22); leaf colours persist across frames.

The Dataset

Data Collection

Phenotyping experiments used 21 Arabidopsis thaliana genotypes grown in an environmentally controlled growth chamber, imaged from above at regular intervals (2 or 4 images per day) with a commercial 4.92 MP RGB camera on an automated gantry designed by Photon Systems Instruments / Scientific Instruments Australia. Each of the 318 plant replicates was recorded as a separate image sequence capturing the full growth trajectory from seedling emergence to a mature rosette. Images are ~530×525 px with the plant centred against a soil background.

Time-series high-throughput data collection of Arabidopsis thalana ecotypes in controlled growth room.

Annotation Protocol

Leaf instance masks were manually annotated in CVAT using a fixed palette of 31 perceptually distinct colours. The first leaf to appear is assigned colour 1 (red), the second colour 2 (blue), and so on by a deterministic ordering. When a new leaf emerges, it receives the next colour in the sequence. The same physical leaf therefore retains the same colour label across all frames in which it appears, so tracking accuracy can be measured directly against the colour-coded ground truth without a Hungarian matching step. All annotations underwent a two-pass quality-control review.

Dataset Statistics

Statistic	Value
Genotypes	21
Plant replicates	318
Total images	17,082
Resolution	~530 × 525
Frames / replicate	53.7
Leaves / image	9.7
Min / max leaves	4 / 26
Growth stages	3
Leaf unique colours	31

Growth stages are defined by leaf count: Early (4–6 leaves, 22.6%), Intermediate (7–10 leaves, 39.8%), and Mature (11+ leaves, 37.6%). Labels are derived automatically by counting unique leaf regions per frame.

22.6%

39.8%

37.6%

EarlyIntermediateMature

Dataset statistics: leaf-count distribution, images per genotype, growth-stage distribution, and the 31-colour annotation palette. — **Figure 2.** Dataset statistics. (a) Leaf-count distribution (mean 9.7, range 4–26). (b) Images per genotype. (c) Growth-stage distribution. (d) The 31-colour annotation palette.

Comparison with Other Public Plant Datasets

PhenoLeaf-TS is the only public top-view RGB phenotyping dataset providing temporally consistent instance labels that enable joint evaluation of segmentation, tracking, and growth stage classification (✓ = supported, – = not).

Dataset	Species	Images	Instance	Time-series	Tracking	Growth	Public
Ara2012	A. thaliana	120	✓	–	–	–	✓
Ara2013	A. thaliana	27	✓	–	–	–	✓
Ara2014	A. thaliana	168	✓	–	–	–	✓
Tobacco	Tobacco	89	✓	–	–	–	✓
MSU-PID	A. thaliana/Bean	2,000+	Partial	✓	–	–	✓
Komatsuna	Komatsuna	900	✓	✓	✓	–	✓
PhenoBench	Sugar beet	1,500+	Panoptic	–	–	–	✓
PhenoLeaf-TS (Ours)	A. thaliana (21)	17,082	✓	✓	✓	✓	✓

Citation

If you use PhenoLeaf-TS, please cite:

@inproceedings{saric2026phenoleafts,
  title     = {PhenoLeaf-TS: A Time-Series Benchmark for Leaf Instance
               Segmentation, Tracking, and Growth Stage Classification},
  author    = {Sari\'c, Rijad and Azam, Basim and Khan, Sarmad and \v{C}ustovi\'c, Edhem},
  booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
  year      = {2026}
}

Acknowledgements

The authors gratefully acknowledge the valuable support and collaboration of Prof. Jim Whelan and Zhejiang University. We also sincerely thank Prof. Mathew G. Lewsey from the Australian Plant Phenomics Network (APPN) node at La Trobe University, as well as Prof. Tony Bacic and Dr. Veronica Borret from the La Trobe Institute for Sustainable Agriculture and Food (LISAF) and the ARC Research Hub for Protected Cropping, La Trobe University, for their guidance, expertise, and continued support throughout this work. We extend our appreciation to Martin Trtilek, Klara Panzarova, Ivan Kashkan, and Sajid Ullah from Photon Systems Instruments for their technical support, collaboration, and assistance with plant phenotyping systems and data-related activities. Their contributions, expertise, and collaborative input have been highly valuable in supporting the development and completion of this research.

This work was supported by the Australian Research Council Industrial Transformation Research Hub for Medicinal Agriculture (ARC MedAg Hub) through Grant Number IH180100006 and the ARC Industrial Transformation Research Hub for Protected Cropping (PC Hub) through Grant Number IH240100024.

Contact

For any inquiries related to the dataset on this website, please contact us via email: