jinyu-xu/TPC-268

---
license: cc-by-nc-sa-4.0
task_categories:
- object-detection
- zero-shot-object-detection
tags:
- counting
- class-agnostic-counting
- few-shot
- plant-phenotyping
- agriculture
- biology
- taxonomy
size_categories:
- 10K<n<100K
---
# 🌳🌻 Plant Taxonomy Meets Plant Counting 🍀🍒 : <br>A Fine-Grained, Taxonomic Dataset for Counting Hundreds of Plant Species
<p align="center">
<a href="https://arxiv.org/abs/2603.21229"><img src="https://img.shields.io/badge/arXiv-2603.21229-b31b1b?style=flat-square&logo=arxiv&logoColor=white" alt="arXiv"></a>
<a href="https://drive.google.com/file/d/1kLlcuyQ1yKE5-TqRkS4CNRgq0VH-RpKc/view?usp=sharing"><img src="https://img.shields.io/badge/Dataset-Download-34A853?style=flat-square&logo=googledrive&logoColor=white" alt="Dataset"></a>
<a href="https://creativecommons.org/licenses/by-nc-sa/4.0/"><img src="https://img.shields.io/badge/License-CC%20BY--NC--SA%204.0-lightgrey?style=flat-square&logo=creativecommons&logoColor=white" alt="License"></a>
<a href="https://github.com/tiny-smart/TPC-268"><img src="https://img.shields.io/badge/GitHub-TPC--268-181717?style=flat-square&logo=github&logoColor=white" alt="GitHub"></a>
</p>

**[CVPR 2026 Oral]** Official repository for the paper **"Plant Taxonomy Meets Plant Counting: A Fine-Grained, Taxonomic Dataset for Counting Hundreds of Plant Species"**.

**Authors:** Jinyu Xu, Tianqi Hu, Xiaonan Hu, Letian Zhou, Songliang Cao, Meng Zhang, Hao Lu*

*TPC-268 is a large-scale dataset for class-agnostic counting (CAC) that explicitly integrates plant taxonomy. It is designed to tackle the unique challenges of visual counting in the natural world, featuring highly diverse and nonrigid plant morphologies.*

## Dataset Overview
* **Images:** 10,000 images spanning extreme observation scales, ranging from canopy-level remote sensing to tissue-level microscopy.
* **Annotations:** 678,050 instance-level points and 30,000 exemplar bounding boxes.
* **Taxonomy:** A complete 7-level Linnaean hierarchy encompassing 2 Kingdoms (Plantae and Fungi), 2 Phyla, 4 Classes, 35 Orders, 83 Families, 192 Genera, and 242 Species.
* **Categories:** 268 fine-grained "Species-Organization" targets, providing precise semantics for counting (e.g., flower, fruit, leaf, stem, stoma, resin).

<p align="center">
<img src="./assets/overview.jpg" width="100%" alt="Sample images from the TPC-268 dataset">
<br>
Sample images from the TPC-268 dataset.
</p>

## Dataset Download

You can download TPC-268 from either of the following links: [Baidu Pan](https://pan.baidu.com/s/1pYET_8I7a6mKmdLYCT0O8g?pwd=jjhv), [Google Drive](https://drive.google.com/file/d/1kLlcuyQ1yKE5-TqRkS4CNRgq0VH-RpKc/view?usp=sharing). Extract the downloaded file into a `TPC-268/` directory. The images are strictly organized by their taxonomy and organization:
`TPC-268/[Genus_Species]/[Organization]/[Genus_Species]_[Organization]_[index].jpg`

```text
TPC-268/
├── Abelmoschus_esculentus/
│ └─ fruit/
│ ├─ Abelmoschus_esculentus_fruit_1.jpg
│ └─ ...
├── Zea_mays/
│ └─ fruit/
│ ├─ Zea_mays_fruit_1.jpg
│ └─ ...
└── ...
```

## Annotations and Splits

Core data files are located in the `annotations/` and `splits/` directories:

* `annotations/tpc268_annotations.json`: Instance-level point annotations and 4-point coordinates for exemplars.
* `annotations/tpc268_taxonomy_ids.json`: Mapping of hierarchical taxonomic levels (Kingdom to Species) to unique numerical IDs.
* `annotations/tpc268_taxonomy_vectors.json`: 7-dimensional taxonomic feature vectors for each plant species.
* `splits/tpc268_[train|val|test].txt`: Lists of relative image paths used for data loading.
* `splits/tpc268_split.json`: Lists of specific species-organization categories partitioned into the train, val, and test sets.

## Scripts

The `tools/` directory provides essential utilities:

**`tpc268_dataset.py`**: A standardized PyTorch `Dataset` class to load images and annotations.
```python
from tools.tpc268_dataset import TPC268Dataset
dataset = TPC268Dataset(data_dir='TPC-268/', split_txt='splits/tpc268_train.txt', anno_json='annotations/tpc268_annotations.json')
```
**`tpc268_visualize_dataset.py`**: Script to overlay annotations on images.
```bash
python tools/tpc268_visualize_dataset.py --img_path TPC-268/Zea_mays/fruit/Zea_mays_fruit_1.jpg --anno_json annotations/tpc268_annotations.json
```
**`TPC268_Annotator.html`**: An HTML tool for browsing and editing annotations directly in the browser.
**`tpc268_generate_benchmark_split.py`**: Script to generate the dataset partitioning lists.

## Benchmark Results

The following table reports the 3-shot counting performance on TPC-268, best performance is highlighted in bold. For other experimental results, please refer to the main paper.

| Method | Backbone | Val MAE | Val RMSE | Val $R^2$ | Test MAE | Test RMSE | Test $R^2$ |
| :---------- | :------- | :-------- | :-------- | :-------- | :-------- | :-------- | :--------- |
| FamNet | R50 | 28.87 | 52.51 | 0.58 | 30.43 | 65.62 | 0.62 |
| BMNet+ | R50 | 29.33 | 77.78 | 0.47 | 27.78 | 57.25 | 0.74 |
| C-DETR | R50 | 22.66 | 77.51 | 0.75 | 22.68 | 57.97 | 0.74 |
| SPDCNet | R18 | 25.66 | 72.49 | 0.52 | 23.70 | 47.53 | 0.64 |
| CountTR | Hybrid | 20.21 | 55.82 | 0.73 | 25.19 | 49.94 | 0.62 |
| SAFECount | R18 | 22.57 | 63.65 | 0.64 | 25.70 | 52.30 | 0.58 |
| LOCA | R50 | 17.26 | 53.19 | 0.75 | **17.51** | **38.37** | **0.78** |
| DAVE | R50 | 16.47 | 52.87 | 0.76 | 17.61 | 40.06 | 0.75 |
| CACVIT | ViT-B | 16.63 | **42.49** | 0.82 | 22.04 | 41.79 | 0.73 |
| CountGD | Swin-B | 18.32 | 54.55 | 0.74 | 19.52 | 50.51 | 0.61 |
| TasselNetV4 | ViT-B | **13.20** | 43.93 | **0.83** | 22.95 | 51.36 | 0.60 |

## Citation

If you find TPC-268 useful for your research, please cite:
```bibtex
@inproceedings{xu2026plant,
title={Plant Taxonomy Meets Plant Counting: A Fine-Grained, Taxonomic Dataset for Counting Hundreds of Plant Species},
author={Xu, Jinyu and Hu, Tianqi and Hu, Xiaonan and Zhou, Letian and Cao, Songliang and Zhang, Meng and Lu, Hao},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
year={2026}
}
```

## License

The TPC-268 dataset is released exclusively for academic research purposes. It is licensed under the [Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0)](LICENSE). Any commercial use, reproduction, or distribution of this dataset without explicit prior consent is strictly prohibited.