Dataset Card for INQUIRE-ReRank

This is a FiftyOne dataset with 16000 samples.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/INQUIRE_Rerank")

# Launch the App
session = fo.launch_app(dataset)

Important Notes:

Although the similarity index over the dataset has already been constructed, you will need to ensure that you have the following installed:

pip install open_clip_torch

After installing this, ensure you downloaded the OpenCLIP and BioCLIP model weights. You can ensure this by running the following lines of code after you have downloaded this dataset:

import fiftyone.zoo as foz

bio_clip_model = foz.load_zoo_model(
    name_or_url="open-clip-torch",
    clip_model="hf-hub:imageomics/bioclip",
    pretrained="",
)

clip_model = foz.load_zoo_model("open-clip-torch")

Then, you can perform a natural language search over the dataset via the App (as shown in the gif above) or via the SDK:

#using the similarity index built with the OpenCLIP model
view_1 = dataset.sort_by_similarity("mongoose on the plains", k=5, brain_key="open_clip_sim")

#using the similarity index built with the BioCLIP model
view_2 = dataset.sort_by_similarity("mongoose on the plains", k=5, brain_key="bio_clip_sim")

To visualize the map, you will need to obtain a free Mapbox Token and set the following environment variable prior to launching the app:

export MAPBOX_TOKEN=XXXXXXXX

Dataset Details

Dataset Description

INQUIRE (A Natural World Text-to-Image Retrieval Benchmark) is a challenging benchmark for evaluating expert-level text-to-image retrieval systems. It consists of 250 expert-level queries paired with comprehensive relevance annotations over a pool of 5 million natural world images (iNat24). Unlike previous text-to-image retrieval benchmarks which contain simple queries with exactly one relevant image per query, INQUIRE includes complex, domain-specific queries with multiple relevant images (ranging from 1-1,500 matches per query, with 33,000 total matches). The benchmark is structured as two tasks: INQUIRE-FULLRANK (full dataset ranking) and INQUIRE-RERANK (reranking of top-100 retrievals). The dataset is designed to bridge the gap between AI capabilities and the needs of real-world scientific inquiry, with a focus on accelerating ecological and biodiversity research.

Curated by: Massachusetts Institute of Technology, University College London, University of Edinburgh, University of Massachusetts Amherst, and iNaturalist
Funded by: Generative AI Laboratory (GAIL) at the University of Edinburgh, the Global Center on AI and Biodiversity Change (NSF OISE-2330423 and NSERC 585136), and the Biome Health Project funded by WWF-UK
Shared by: The INQUIRE research team
Language(s) (NLP): en
License: CC BY-NC 4.0

Dataset Sources

Repository: https://github.com/inquire-benchmark/INQUIRE
Project Page: https://inquire-benchmark.github.io/
Paper: Vendrow, E., Pantazis, O., Shepard, A., Brostow, G., Jones, K. E., Mac Aodha, O., Beery, S., & Van Horn, G. (2024). INQUIRE: A Natural World Text-to-Image Retrieval Benchmark. arXiv:2411.02537v3
Dataset on HF: https://huggingface.co/datasets/evendrow/INQUIRE-Rerank

To see the code for how to parse this dataset into FiftyOne format, checkout this notebook on Colab.

Uses

Direct Use

This dataset is intended to be used for:

Benchmarking text-to-image retrieval models on expert-level queries
Evaluating reranking methods for image retrieval
Developing more effective retrieval systems for ecological and biodiversity research
Testing multimodal models' understanding of fine-grained natural world concepts
Research into domain-specific retrieval that requires expert knowledge

Out-of-Scope Use

This dataset should not be used for:

Making conservation assessments based on predictions from biased or inaccurate models
Commercial applications (prohibited by the CC BY-NC 4.0 license)
Training general-purpose image generation models
Applications that might create inaccurate or harmful associations between queries and wildlife images
Creating derivative datasets that do not respect the original image licenses

Dataset Structure

The INQUIRE benchmark consists of two main components:

iNat24 Dataset:
- 5 million natural world images spanning 10,000 species
- Collected from the iNaturalist platform between 2021-2023
- Each image includes metadata: species label, location, observation time, license information, and full taxonomic classification
INQUIRE Queries and Annotations:
- 250 expert-level text queries
- Each query belongs to one of four supercategories (Appearance, Behavior, Context, or Species)
- Further subdivided into 16 fine-grained categories (e.g., Health and Disease, Feeding and Hydration)
- Comprehensive relevance annotations for each query (33,000 total positive matches)
- Split into 50 validation and 200 test queries
INQUIRE-RERANK Task:
- Fixed set of 100 images per query (retrieved using CLIP ViT-H-14)
- 200 queries filtered to ensure at least one positive image is among the top 100
- Split into 40 validation and 160 test queries

Dataset Creation

Curation Rationale

INQUIRE was created to address several limitations in existing text-to-image retrieval benchmarks:

Current benchmarks are small (1,000-5,000 images)
They contain simple queries and are no longer challenging for modern models
Each query typically has only one relevant image, unlike real-world retrieval scenarios
They lack expert-level concepts that would require domain knowledge

The natural world domain was chosen because it contains rich visual concepts requiring expert knowledge, and improvements in retrieval could accelerate scientific discovery in ecology and biodiversity research.

Source Data

Data Collection and Processing

iNat24 was created from an iNaturalist observation database export (generated on 2023-12-30)
Images were filtered to include only observations from 2021-2023
The dataset uses the iNat21 taxonomy (for compatibility with that dataset)
Images were selected using a spatial-temporal clustering approach to reduce geographic bias
All images were resized to have a maximum dimension of 500px
Images containing human faces, personally identifiable information, or inappropriate content were filtered out

Who are the source data producers?

The images in iNat24 were contributed by citizen scientists on the iNaturalist platform. These are typically nature enthusiasts who photograph and document plants, animals, and other organisms they encounter, and upload them to the platform with species identification.

Annotations

Annotation process

Queries were collected through interviews with experts across ecological and environmental domains
Only queries that could be discerned from images alone and feasibly labeled over the entire dataset were retained
For each query, annotators labeled candidate images as either relevant or not relevant
When applicable, species labels were used to narrow down the search space for comprehensive labeling
For queries that couldn't be filtered by species, annotators labeled top CLIP retrievals until at least 100 consecutive retrievals were not relevant
In total, 194,334 images were labeled, with 32,696 identified as relevant across all queries

Who are the annotators?

The annotations were performed by a carefully selected team of paid MSc students or equivalent, many with expertise in ecology allowing for labeling of difficult queries. Annotators were paid at the equivalent of $15.50 per hour.

Personal and Sensitive Information

The dataset was filtered to remove images containing human faces, personally identifiable information, or other sensitive content. A combination of automated methods (RetinaFace face detection) and manual inspection was used to identify and remove such images.

Bias, Risks, and Limitations

Geographic bias: The images in iNat24 are not uniformly distributed globally, with higher representation from North America, Europe, and parts of Australasia, reflecting the spatial biases present in the iNaturalist platform.
Taxonomic bias: Some species groups may be overrepresented compared to others based on popularity among citizen scientists.
Potential for misinterpretation: The expert-level nature of some queries means that models might make confident but incorrect predictions.
Limited to visual information: The dataset only captures what can be observed in images, which may not capture all ecologically relevant information.
Seasonal bias: Observations might be skewed toward certain seasons when volunteer activity is higher.

Recommendations

Users should be aware of the geographic biases when applying models trained or evaluated on this data to regions with limited representation.
Conservation or scientific decisions should not be made solely based on model predictions without expert validation.
When developing retrieval systems, consider implementing filters to prevent potentially harmful text-to-image associations.
Future work could focus on addressing the geographic and taxonomic biases through targeted data collection.
Systems evaluated on this benchmark should be transparent about their limitations when deployed in real-world scientific applications.

Citation

BibTeX:

@inproceedings{vendrow2024inquire,
  title={INQUIRE: A Natural World Text-to-Image Retrieval Benchmark},
  author={Vendrow, Edward and Pantazis, Omiros and Shepard, Alexander and Brostow, Gabriel and Jones, Kate E. and Mac Aodha, Oisin and Beery, Sara and Van Horn, Grant},
  booktitle={38th Conference on Neural Information Processing Systems (NeurIPS 2024) Track on Datasets and Benchmarks},
  year={2024},
  eprint={2411.02537},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}

APA: Vendrow, E., Pantazis, O., Shepard, A., Brostow, G., Jones, K. E., Mac Aodha, O., Beery, S., & Van Horn, G. (2024). INQUIRE: A Natural World Text-to-Image Retrieval Benchmark. In 38th Conference on Neural Information Processing Systems (NeurIPS 2024) Track on Datasets and Benchmarks.

Glossary

mAP@50: Mean Average Precision at 50, a metric for evaluating retrieval quality
nDCG: Normalized Discounted Cumulative Gain, a ranking metric that considers the relative ordering of retrieved items
MRR: Mean Reciprocal Rank, a measure of the rank of the first correct retrieval
Reranking: The process of refining an initial set of retrievals using a more sophisticated model
iNaturalist: A citizen science platform where users contribute observations of biodiversity

Dataset Card Authors

The INQUIRE research team: Edward Vendrow, Omiros Pantazis, Alexander Shepard, Gabriel Brostow, Kate E. Jones, Oisin Mac Aodha, Sara Beery, and Grant Van Horn

Dataset Card Contact

For questions or issues related to the dataset, please contact the team through the GitHub repository: https://github.com/inquire-benchmark/INQUIRE

image imagewidth (px) 250 1.48k