Bio-VQA

Bio-VQA is a biological visual question answering benchmark of 1.86M questions over 217K images from 11 public datasets, in four components:

Component	Tasks	Images	Questions	Image source
C1 Hierarchical taxonomic reasoning	9	100,000	890,000	TreeOfLife-200M, referenced by uuid
C2 Fine-grained attribute grounding	4	23,576	824,104	CUB-200-2011 originals referenced by path; marker renders included
C3 Quantitative spatial reasoning	8	78,139	132,504	included
C4 Individual re-identification	3	15,000	15,000	included (side-by-side pair images)

Paper: How Finely Can Vision–Language Models See Biology? Bio-VQA: A Benchmark for Taxonomic, Morphological, Spatial and Individual Reasoning, WACV 2027. Code: https://github.com/Imageomics/Bio-VQA

Layout

lite/          Bio-VQA Lite: 500 questions per task, 12,000 in total
C1/            question files, one per organismal group
C2/            QA.csv and the marker-localization renders (zip shards)
C3/            one question file and one image archive per task
C4/            one question file and one image archive per species and condition
predictions/   per-item answers of all 18 models and both specialists, as scored in the paper

Each question file has the question, the gold answer and the image reference. In C2–C4, image_zip names the archive and image_in_zip the file inside it.

Images

• C1 images come from TreeOfLife-200M. Each row carries the TreeOfLife-200M uuid; 1.Hierarchical_taxonomic_reasoning/Tree-of-Life-200M/download.py in the code repository fetches them.
• C2 original photographs come from CUB-200-2011 (download from the official Caltech release; paths in filename are relative to its root). The 11,788 marker-localization images are derived renders and are included here.
• C3 and C4 images are included. C4 pair images are renamed to neutral sequential names.

Licenses

The benchmark annotations are released under CC BY-NC 4.0. Images keep the license of their source dataset: TreeOfLife-200M (per-image licenses), CUB-200-2011 (non-commercial research use), WAID, Boxes on Bees, 2018-NEON-Beetles, DePondFi, WhaleSharkID, BelugaID, HyenaID, LeopardID and GZCD.

Citation

@inproceedings{rayeed2027biovqa,
  title     = {How Finely Can Vision--Language Models See Biology? {Bio-VQA}: A Benchmark for Taxonomic, Morphological, Spatial and Individual Reasoning},
  author    = {Rayeed, S M and Ma, Qian and Stewart, Charles V. and Ma, Yao},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year      = {2027}
}