AgroGround

Annotations, benchmark, prediction files and contamination flags of the paper "AgroGround: Multi-Granularity Grounded Recognition in Agriculture" (Abdulla Alshehhi, Zongyan Han, Rao Anwer; under review).

• Paper: https://arxiv.org/abs/2610.04425
• Code: https://github.com/AB-Abdulla/AgroGround
• Canonical versioned record: Zenodo, https://doi.org/10.5281/zenodo.23149180 — this Hugging Face repository is a mirror for convenience and discoverability. If the two ever differ, the Zenodo record is authoritative.

No source images and no question-answer text are included. AgroGround is an annotations-only release: every row is keyed to its source dataset by three fields — dataset (one of cddm, leafnet, leafbench, mirage, agromind, agrocot, agrobench, agmmu), sample_id (the source's own sample identifier) and image_id (the image's path relative to the source's release root, e.g. cddm/images/Blueberry,Healthy/plant_64407.jpg). Users obtain the images from the original dataset releases and join them by image_id. Boxes are pixel coordinates [x1, y1, x2, y2] of the original image; image width and height are stored alongside.

What is here

Config / path	Contents
annotations	The 794,850 training examples (391,758 lesion-level, 403,092 object-level) with label, kind, crop, crop-aware prompt, granularity, boxes and the 80/10/10 split, in eight per-source shards
healthy_negatives	The 43,000 healthy images used as negatives (used_in marks the 12% set and the 21,500-image 6% subset)
benchmark_v4	The human-verified benchmark, 1,480 images (508 lesion-level, 467 object-level, 505 healthy; 3,463 boxes) — the version the submitted tables were computed on
benchmark_v4_1	Benchmark v4.1, 1,478 images — withdraws the two images listed in benchmark/v4.1_exclusions.json (copies of healthy training negatives found by the released audit script)
recognition_manifest_v4, recognition_manifest_v4_1	Candidate lists for the recognition protocol
contamination_flags	One row per training image file (231,323): dataset, source split, AgroGround splits, negative use, benchmark membership — use it to exclude images before evaluating a model trained on AgroGround on any of the eight sources
predictions/ (plain files)	The 63 prediction files of every model and protocol in the paper, keyed by coco_id; evaluation/score.py in the code repository reproduces the paper's tables from them. The ten PlantSeg files carry predictions only (PlantSeg's ground truth is not redistributed)
benchmark/gold_taxonomy_prompts_v4.json	The annotation-time prompt per benchmark image (taxonomy-prompt teacher baseline)
benchmark/benchmark_healthy_lock_rounds1_2.json	The 430 healthy benchmark images excluded from the negative pool
phash_index.json	64-bit perceptual hash per image_id for the 171,580 training files and the 1,480 benchmark images (the disjointness audit's input)
LICENSES.md, MANIFEST.json	Per-shard licenses; file sizes and row counts

Reproducing the paper

from datasets import load_dataset
bench = load_dataset("abdulla90/AgroGround", "benchmark_v4")

Every number in the paper is reproducible from these files with the scoring script in the code repository: place the contents of predictions/ and benchmark/ in one folder and run python evaluation/score.py --pred_dir <folder>.

Licenses

Each annotation shard carries the license of its source dataset (see LICENSES.md and the LICENSE.txt inside each shard): CC BY-SA 4.0 for the MIRAGE, AgroCoT and AgMMU shards; CC BY-NC 4.0 for the LeafNet and LeafBench shards; CC BY 4.0 for the CDDM, AgroMind and AgroBench shards. The evaluation set is partitioned by source under the same terms. The prediction files, flags, hash index and exclusion list are CC BY 4.0. Please cite the source datasets when using the corresponding shards.

Citation

@misc{alshehhi2026agroground,
  title={AgroGround: Multi-Granularity Grounded Recognition in Agriculture},
  author={Alshehhi, Abdulla and Han, Zongyan and Anwer, Rao},
  year={2026},
  eprint={2610.04425},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.04425}
}