README.md

---
license:
- cc-by-4.0
- cc-by-nc-4.0
pretty_name: VTOS Benchmarks (LVIS-Count, PlantSeg-OOD)
task_categories:
- object-detection
- image-segmentation
- visual-question-answering
size_categories:
- n<1K
---

# VTOS Benchmarks

Two small benchmarks used in [VTOS: Learning to Orchestrate Vision Tools by Co-Searching Solutions
and Observers](https://arxiv.org/abs/2606.20728) (EMNLP 2026). Code: https://github.com/jinchaogjc/VTOS.
Each benchmark has a 60 / 20 / 100 train / val / test split.

| Subset | Task | Items | Images | Source |
|---|---|---|---|---|
| `lvis_count/` | object counting | 180 | 180 | LVIS (images from COCO) |
| `plantseg_ood/` | plant-disease segmentation | 180 | 180 | PlantSeg |

## Layout

```
lvis_count/
  benchmark_train.json   benchmark_val.json   benchmark_test.json
  images/lvis_count_001.jpg ... lvis_count_180.jpg
plantseg_ood/
  benchmark_train.json   benchmark_val.json   benchmark_test.json
  images/<plant>_<disease>_<id>.jpg
```

Each JSON file is a list of items. `image_path` is relative to the subset's `images/` directory.
Every image is referenced by exactly one item.

## Fields

Coordinates are normalized to [0, 1]. Bounding boxes are `[x_top_left, y_top_left, width, height]`
(FiftyOne convention).

**lvis_count**

| Field | Description |
|---|---|
| `task_id` | item identifier |
| `image_path` | file name under `images/` |
| `question` | counting question |
| `target_class` | LVIS category to count |
| `count` | ground-truth count (= number of boxes) |
| `density_tier` | density bucket |
| `bounding_boxes` | one box per target instance |
| `meta_info.source_image_id` | sample id in the FiftyOne LVIS export |
| `meta_info.source_filepath` | file path in that export; the file name is the COCO image id |

**plantseg_ood**

| Field | Description |
|---|---|
| `task_id` | item identifier |
| `image_path` | file name under `images/` |
| `question` | localization question |
| `plant` | host plant |
| `target_class` | disease name |
| `mask_ratio` | fraction of image area covered by the disease mask |
| `difficulty_tier` | difficulty bucket |
| `bounding_boxes` | one box per region |
| `segmentations` | one polygon per region, list of `[x, y]` points |
| `meta_info` | `n_bboxes`, `n_polygons`, `source_filepath` in the FiftyOne PlantSeg export |

## Sources and licenses

| Subset | License | Images |
|---|---|---|
| `lvis_count/` | CC BY 4.0 | COCO images; copyright remains with the image owners, use subject to the Flickr Terms of Use |
| `plantseg_ood/` | CC BY-NC 4.0 (non-commercial use only) | from PlantSeg, same license |

**lvis_count.** Annotations are taken from LVIS, licensed CC BY 4.0
(https://www.lvisdataset.org/dataset). LVIS images come from COCO. The COCO Consortium does not
own the image copyrights; use of the images must follow the Flickr Terms of Use
(https://cocodataset.org/#termsofuse). Per-image Flickr licenses were not verified for this subset.
Local copy obtained from the FiftyOne LVIS export. Image files were renamed to `lvis_count_XXX.jpg`;
image content is unmodified.

**plantseg_ood.** Taken from PlantSeg (Wei et al., 2024). Zenodo lists v1 (2024-08,
https://doi.org/10.5281/zenodo.13293891) as CC BY-NC-ND 4.0 and the latest version, v7
(2025-11, https://zenodo.org/records/17719108), as CC BY-NC 4.0. This subset is distributed under
CC BY-NC 4.0 and is for non-commercial use only. Local copy obtained from `Voxel51/PlantSeg-Test`
on Hugging Face; which Zenodo version that export is based on could not be verified.
Bounding boxes match the FiftyOne export for 177 of 180 items. GPS metadata was removed from three
images; pixel data of all images is unmodified.

## Citation

If you use these benchmarks, please cite:

```bibtex
@misc{ge2026vtos,
  title={{VTOS}: Learning to Orchestrate Vision Tools by Co-Searching Solutions and Observers},
  author={Ge, Jinchao and Liu, Lingqiao and Zhao, Shuwen and Wang, Lei},
  year={2026},
  eprint={2606.20728},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2606.20728}
}
```