S M Rayeed
Home / Research

Biological visual question answering

General-purpose VLMs are increasingly used to read biological images. BioVQA asks whether they can: it separates taxonomic knowledge, fine-grained attribute grounding, quantitative spatial reasoning and individual identity, and scores every model against an image-blind floor and a specialist ceiling.

Project page1.87M questions216K images18 open-weight VLMs, 3.75B to 38.4BHierarchical taxonomic reasoningTreeOfLife-200M, 10 organism groupsQ: Which genus does this organism belong to?Fine-grained attribute groundingCUB-200-2011, 824K questionsQ: What color is the marked part?Quantitative spatial reasoningCounting, OCR, distance, viewpointQ: How many specimens are on this tray?Individual re-identification5 species, 3 conditions, scored with d′Q: Are these the same individual?kingdom to species,checked against GBIFsynonyms; BioCLIP-2as specialist ceilingCAGS metric withlaterality checks;three-annotatorhuman validationOCRdistancesameor not?
WACV 2027First author

BioVQA: A Benchmark for Biological Question Answering across Taxonomic, Fine-grained and Spatial Reasoning

S M Rayeed et al.
IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2027), Datasets and Benchmarks track

1.87M questions over 216K biological images, testing 18 open-weight VLMs on four abilities: hierarchical taxonomic reasoning, fine-grained attribute grounding, quantitative spatial reasoning, and individual re-identification. Every model is placed between an image-blind floor and a specialist ceiling.

Project pagePaperCode

Other research areas

Fine-grained taxonomic identificationKnowledge-based VQA and retrievalAI-ready biodiversity dataAnimal re-identificationSign language recognition