InSpect Dataset

On this page

InSpect Overview

InSpect overview with three panels: (a) collection images and labels; (b) aligned specimen crops, hierarchical taxonomy, structured metadata, and spatial-aware and singular part masks; (c) open taxonomic recognition and anatomical part segmentation benchmarks.
Figure 2: Overview of InSpect and its two benchmarks. (a) The dataset contains 48,184 natural-history collection images spanning 653 taxa, with specimens and their accompanying collection labels. (b) Each specimen is associated with an aligned crop, hierarchical taxonomy, and label-derived structured metadata. A subset of 1,859 images provides fine-grained anatomical part masks in two forms: ① spatial-aware and ② singular. (c) InSpect supports open taxonomic recognition from specimen crops with optional metadata under zero-shot, fine-tuned, and unseen-taxon settings, as well as anatomical part segmentation under supervised and text-guided settings.

Data Structure

The release is hosted on Hugging Face. Images and anatomical annotations are distributed as ZIP archives; specimen metadata is provided in CSV and JSONL formats.

InSpect/
├── README.md
├── original_images.zip
├── final_benchmark_data_crops.zip
├── specimen_benchmark_metadata.csv
├── specimen_benchmark_metadata.jsonl
├── segmentation_images.zip
└── segmentation_annotations.zip
FileSizeContents
original_images.zip248.15 GBOriginal collection photographs, including specimens and their physical labels.
final_benchmark_data_crops.zip17.50 GBCropped insect images for visual recognition; use these to exclude physical label text from the visual input.
specimen_benchmark_metadata.csv32.07 MBOne row per specimen; image identifiers, taxonomy, benchmark splits and label-derived context.
specimen_benchmark_metadata.jsonl57.80 MBJSONL version of the specimen metadata.
segmentation_images.zip801.91 MBImages for the anatomical part segmentation subset.
segmentation_annotations.zip13.96 MBCOCO-style JSON annotations for singular and spatial-aware part labels.

File sizes checked on 30 October 2026; decimal units (1 GB = 1,000,000,000 bytes). Download only the files needed for your task.

Recognition: start with final_benchmark_data_crops.zip and either metadata file. Part segmentation: use segmentation_images.zip with segmentation_annotations.zip. Original full images are available separately for studying collection images and labels.

Metadata fields and annotation files

The released CSV contains 48,184 records and 31 columns. Key identifiers are image_name and catalog_number; taxonomy is stored in order, family, genus, species and taxon_key. The split column uses train, val, test_seen and test_unseen. Context fields include locality, collection date, collector, life stage and sex, with missing values preserved.

For metadata-based recognition, exclude direct taxonomic fields as described in the paper. The annotation archive contains the following files (filenames, including spaces, are shown as released):

annotations/
├── singular_sam3_train.json
├── singular_sam3_val.json
├── singular_sam3_test.json
├── sam3_train 1.json
├── sam3_val 1.json
└── sam3_test 1.json

Each JSON file contains COCO-style images, annotations and categories entries. The singular files contain 11 singular anatomical labels; the other three files use a 31-category vocabulary with spatial-aware distinctions.

Annotations Examples

Recognition

Explore a specimen’s original collection photograph, aligned insect crop, hierarchical taxonomy and label-derived metadata. Select a specimen below to inspect the corresponding record.

Open image for closer inspection ↗

Segmentation

Explore anatomical part annotations from InSpect. Select a specimen, switch between singular and spatial-aware labels, or adjust the mask opacity to inspect the underlying image. These overlays show released ground-truth annotations, not model predictions.

Tools

Code: Recognition and Segmentation

Contribute

We welcome contributions of the following:

  • Published benchmarking results of algorithms applied to the InSpect dataset.
  • Corrections to existing descriptions or ground truth.
  • Other aspects of the InSpect.

Please email junhaozhang.ai@gmail.com if you have any questions about contributing datasets.

Citation

If you find our InSpect dataset is useful in your research, please consider cite:

Provisional BibTeX; replace with the official export when available.

@misc{zhang2026inspect,
  title = {InSpect: A Curated Natural History Collection Dataset
           for Insect Specimen Understanding},
  author = {Zhang, Junhao and Xu, Feng and Ravendran, Ahalya
            and Fisher, Nicole and Tidd, Brendan
            and Petersson, Lars and Wang, Dadong and Li, Xun},
  year = {2026},
  note = {NeurIPS 2026; provisional metadata},
  url = {https://openreview.net/forum?id=51J0UrBZZy}
}

  Download .bib

Copyright:

Our dataset follows the copyright Creative Commons BY-NC-ND 4.0 license.

Search