Dataset Card for VLM4Bio

Dataset Details

VLM4Bio is a benchmark dataset of scientific question-answer pairs used to evaluate pretrained VLMs for trait discovery from biological images. VLM4Bio consists of images of three taxonomic groups of organisms: fish, birds, and butterflies, each containing around 10k images.

Dataset Description

VLM4Bio is a large, annotated dataset, consisting of 469K question-answer pairs involving around 30K images from three groups of organisms: fish, birds, and butterflies, covering five biologically relevant tasks. The scientifically relevant tasks in organismal biology includes species classification, trait identification, trait grounding, trait referring, and trait counting. These tasks are designed to test different facets of VLM performance in organismal biology, ranging from measuring predictive accuracy to assessing their ability to reason about their predictions using visual cues of known biological traits. For example, the tasks of species classification test the ability of VLMs to discriminate between species, while in trait grounding and referring, we specifically test if VLMs are able to localize morphological traits (e.g., the presence of fins of fish or patterns and colors of birds) within the image. We consider two types of questions in this dataset. First, we consider open-ended questions, where we do not provide any answer choices (or options) to the VLM in the input prompt. The second type is multiple-choice (MC) questions, where we provide four choices of candidate answers for the VLM to choose from (out of which only one is correct while the remaining three are randomly selected from the set of all possible answers).

Supported Tasks and Leaderboards

The following figure illustrates VLM4Bio tasks with different question types.

The following table demonstrates the leaderboard of the VLM baselines in terms of zero-shot accuracy.

Languages

English, Latin

Dataset Structure

Instructions for downloading the dataset

Data Instances

Data Fields

Data Splits

Curation Rationale

Source Data

Data Collection and Processing

We collected images of three taxonomic groups of organisms: fish, birds, and butterflies, each containing around 10k images. Images for fish (Fish-10k) were curated from the larger image collection, FishAIR, which contains images from the Great Lakes Invasive Network Project (GLIN) and Integrated Digitized Biocollections (iDigBio). These images originate from various museum collections such as INHS, FMNH, OSUM, JFBM, UMMZ and UWZM. We created the Fish-10k dataset by randomly sampling 10K images and preprocessing the images to crop and remove the background. For consistency, we leverage GroundingDINO to crop the fish body from the background and Segment Anything Model (SAM) to remove the background. We curated the images for butterflies (Butterfly-10k) from the Jiggins Heliconius Collection dataset, which has images collected from various sources. We carefully sampled 10K images for Butterfly-10K from the entire collection to ensure the images capture unique specimens and represent a diverse set of species by adopting the following two steps. First, we filter out images with more than one image from the same view (i.e., dorsal or ventral). Second, we ensure each species has a minimum of 20 images and no more than 2,000 images. The images for birds (Bird-10k) are obtained from the CUB-200-2011 dataset by taking 190 species for which the common name to scientific name mapping is available. This results in a fairly balanced dataset with around 11K images in total.

Annotations

The scientific names for the images of Fish-10k and Butterfly-10k were obtained directly from their respective sources. For Bird-10k, we obtained the scientific names from the iNatLoc500 dataset. We curated around 31K question-answer pairs in both open and multiple-choice (MC) question-formats for evaluating species classification tasks. The species-level trait presence/absence matrix for Fish-10k was manually curated with the help of biological experts co-authored in this paper. We leveraged the Phenoscape knowledge base with manual annotations to procure the presence-absence trait matrix. For Bird-10k, we obtained the trait matrix from the attribute annotations provided along with CUB-200-2011. We constructed approximately 380K question-answer pairs for trait identification tasks. For grounding and referring VQA tasks, the ground truths were manually annotated with the help of expert biologists on our team. We manually annotated bounding boxes corresponding to the traits of 500 fish specimens and 500 bird specimens, which are subsets of the larger Fish-10k and Bird-10k datasets, respectively. We used the CVAT tool for annotation.

Personal and Sensitive Information

None

Considerations for Using the Data

The fish-10 K and Butterfly-10K datasets are not balanced for the species classification task, while the bird-10 K dataset is balanced. Since the fish images are collected from different museums, they may inherit a small bias.

Licensing Information

The data (images) contain a variety of licensing restrictions ranging from CC0 to CC BY-NC. Each image in this dataset is provided under the least restrictive terms allowed by its licensing requirements as provided to us (i.e., we impose no additional restrictions past those specified by licenses in the license file).

This dataset (the compilation) has been licensed under CC BY 4.0. However, images may be licensed under different terms (as noted above). For license and citation information by image, see our license file.

Citation

Please cite the following:

@misc{maruf2024vlm4bio,
      title={VLM4Bio: A Benchmark Dataset to Evaluate Pretrained Vision-Language Models for Trait Discovery from Biological Images}, 
      author={M. Maruf and Arka Daw and Kazi Sajeed Mehrab and Harish Babu Manogaran and Abhilash Neog and Medha Sawhney and Mridul Khurana and James P. Balhoff and Yasin Bakis and Bahadir Altintas and Matthew J. Thompson and Elizabeth G. Campolongo and Josef C. Uyeda and Hilmar Lapp and Henry L. Bart and Paula M. Mabee and Yu Su and Wei-Lun Chao and Charles Stewart and Tanya Berger-Wolf and Wasila Dahdul and Anuj Karpatne},
      year={2024},
      eprint={2408.16176},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2408.16176}, 
}

Acknowledgements

This work was supported by the Imageomics Institute, which is funded by the US National Science Foundation's Harnessing the Data Revolution (HDR) program under Award #2118240 (Imageomics: A New Frontier of Biological Information Powered by Knowledge-Guided Machine Learning). Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.

Dataset Card Authors

M. Maruf (Contact mail: [email protected])