AgroVG

AgroVG is a large-scale benchmark for agricultural visual grounding. It contains two evaluation tasks:

• T1: box-set visual grounding, where a model receives an agricultural referring expression and returns a set of bounding boxes.
• T2: query-level mask grounding, where a model receives an agricultural referring expression and returns a binary segmentation mask.

This repository is the anonymized public-release package prepared for peer review and archival hosting. It contains only release data and metadata; construction scripts, model predictions, and evaluation code should be hosted in a separate code repository.

Dataset URL for review: https://huggingface.co/datasets/sauryrs/AgroVG

Dataset Composition

Task	Images	Queries	Output target	Main files
T1	6,526	6,526	Bounding-box set	t1/annotations, t1/images, t1/queries
T2	3,545	3,545	Query-level binary mask	t2/annotations, t2/images, t2/instance_maps, t2/queries
Total	10,071	10,071	Boxes or masks	t1, t2

Repository Structure

AgroVG/
  README.md
  LICENSE
  CITATION.cff
  croissant.json
  NOTICE
  source_licenses.json
  t1/
    annotations/
      annotations.jsonl
      dev.jsonl
      test.jsonl
      final_audit_table.csv
      policy.json
      final_stats.json
      split_stats.json
      kept_ids.txt
      dropped_ids.txt
    images/
      <source>/
    queries/
      queries_dev.jsonl
      queries_test.jsonl
      query_policy.json
      query_stats.json
  t2/
    annotations/
      annotations.jsonl
      dev.jsonl
      test.jsonl
      final_audit_table.csv
      policy.json
      final_stats.json
      split_stats.json
      kept_ids.txt
      dropped_ids.txt
    images/
      <source>/
    instance_maps/
      <source>/
    queries/
      queries_dev.jsonl
      queries_test.jsonl
      query_policy.json
      query_stats.json

All rgb_relpath and mask_relpath fields are relative to the corresponding task root (t1/ or t2/).

Annotation Records

Each image annotation record contains:

• image_id: stable image identifier.
• source: normalized source-dataset key.
• scene_family: benchmark-level scene family.
• sensor_type: imaging context when available.
• task_candidates: task list, e.g. ["T1"] or ["T2"].
• annotation_type: bbox for T1 and instance_mask for T2.
• rgb_relpath: relative RGB image path.
• mask_relpath: relative uint16 instance-map path for T2.
• width, height: image dimensions.
• group_id: group identifier used to avoid split leakage.
• split_source: source-side split label when available.
• benchmark_split: dev or test.
• instances: normalized instance annotations.
• meta: source-specific metadata retained for auditability.

T1 instances include object_id, class_family_global, source-derived class labels, and bbox_xyxy. T2 instances additionally include instance_local_id, mask-derived geometry, mask area statistics, and mask-generation provenance when applicable.

Query Records

Each query record contains the common fields:

• query_id
• image_id
• benchmark_split
• source
• scene_family
• sensor_type
• rgb_relpath
• width
• height
• group_id
• query
• query_type
• program_type
• template_id
• target_object_ids
• target_count
• is_empty
• meta

T1 query records additionally include target_boxes. T2 query records additionally include mask_relpath and target_instance_local_ids.

Splits

AgroVG uses dev and test splits. Split construction is group-aware: records sharing a group_id are assigned to the same benchmark split to reduce leakage across visually related images.

Evaluation

T1 evaluates set-level box grounding with IoU-thresholded matching. T2 evaluates query-level mask grounding with overlap-based mask metrics and separate target-absent accuracy. Evaluation scripts are not included in this data-only package and should be released in the accompanying code repository.

Licensing

AgroVG-specific annotations, query templates, audit metadata, and derived benchmark metadata are released under Creative Commons Attribution 4.0 International (CC BY 4.0), included in LICENSE.

Images and source-derived annotations retain the licenses, terms, and redistribution permissions of their original source datasets. Source-level license and attribution information is summarized in source_licenses.json and NOTICE. Non-public permission correspondence is intentionally not included in this anonymized review package.

Responsible Use and Limitations

AgroVG is intended for research on agricultural visual grounding, referring-expression comprehension, referring segmentation, and robustness analysis across agricultural domains. It is not intended for direct deployment in farm-management, pesticide, disease-treatment, yield-estimation, or safety-critical agricultural decision systems without additional domain validation.

Known limitations include source-dataset heterogeneity, uneven geographic and crop coverage, possible source-specific annotation biases, and the fact that AgroVG normalizes existing source labels rather than re-certifying all taxonomic labels from scratch.

Citation

This anonymized release is associated with a NeurIPS 2026 submission. Please cite the accompanying anonymous submission during review. The citation metadata in CITATION.cff should be updated with final author names, venue information, and DOI after the review process.