File size: 2,419 Bytes
16826a0
 
ce9ea07
16826a0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ce9ea07
16826a0
ce9ea07
16826a0
 
 
 
 
 
 
 
ce9ea07
16826a0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
---
license: cc-by-nc-4.0
pretty_name: InSpect
task_categories:
- image-classification
- zero-shot-image-classification
- image-segmentation
tags:
- insects
- biodiversity
- natural-history-collections
- taxonomy
- open-vocabulary-recognition
- segmentation
- anatomical-part-segmentation
configs:
- config_name: metadata
  data_files:
  - split: full
    path: specimen_benchmark_metadata.csv
---

# InSpect

InSpect is a curated natural history collection dataset for visual insect specimen understanding. It contains digitized insect specimen images with aligned crops, hierarchical taxonomy, label-derived structured metadata, and fine-grained anatomical part annotations.

## Files

- `specimen_benchmark_metadata.csv`: main metadata table. Each row corresponds to one specimen image and includes split information, taxonomic labels, image/crop paths, and label-derived structured metadata.
- `specimen_benchmark_metadata.jsonl`: JSONL version of the metadata table.
- `final_benchmark_data_crops.zip`: cropped insect images used for recognition and segmentation.
- `segmentation_images.zip`: images used for the anatomical part segmentation subset.
- `segmentation_annotations.zip`: COCO-style anatomical part segmentation annotations.
- `original_images.zip`: The original images before crop. Most of the images contains on insect and the noisy metadata.

## Dataset Structure

The main metadata file contains the structured records used by the benchmark. The image files and segmentation annotations are provided as compressed archives for download.

Recognition experiments use cropped insect images rather than original full images to avoid direct reading of physical specimen labels. Metadata fields that directly encode taxonomy are removed when evaluating metadata-based recognition.

## Benchmark Tasks

### Open Taxonomic Recognition

Models predict taxon-level labels from cropped insect images. The benchmark includes zero-shot, fine-tuned, and unseen-taxon evaluation, with additional metadata-based settings for probing weak specimen context.

### Fine-grained Anatomical Part Segmentation

Models segment insect anatomical structures such as antennae, legs, wings, head, thorax, and abdomen. The benchmark includes supervised and text-guided open-vocabulary segmentation settings.

## License

This dataset is released for non-commercial research and educational use under the CC BY-NC 4.0 license.