IBBI-bio/bark-ambrosia-beetle-benchmark

Bark and Ambrosia Beetle Detection Benchmark

Version 2.0.1 · 14,491 images · 175 species · 21 tribes · 70 genera · COCO detection format

A specimen-disjoint, species-level object-detection benchmark for bark and ambrosia beetles (Coleoptera: Curculionidae: Scolytinae and Platypodinae), derived from the Bark and Ambrosia Gallery (https://barkandambrosiagallery.org/). Species determinations are made or reviewed by taxonomists; individual specimens carry bounding boxes.

This Zenodo record is the archival copy of v2.0.1. The same version is mirrored at https://huggingface.co/datasets/IBBI-bio/bark-ambrosia-beetle-benchmark. The two are byte-identical for the annotation files; cite the DOI of this record for reproducible work.

Licensing note. Images carry mixed Creative Commons licences and there is no single dataset-level licence. The Zenodo licence field is set to "Other (Open)". The authoritative, per-image licence record is image_licences.csv in this record. See Licensing below before any redistribution or commercial use.

---

Why this benchmark exists

Bark and ambrosia beetles are among the most damaging invasive forest insects worldwide, and the regulatory decisions they trigger — quarantine, eradication, port interception — are made at the level of species. Most wood-boring insects intercepted in wood packaging material at United States ports are never identified below family, because the specialists who can separate morphologically near-identical congeners are few.

Automated identification is the obvious remedy, but the four things a model needs have not previously co-occurred in one dataset: expert determination, per-specimen localization, coverage of how these beetles are actually photographed, and a disjoint evaluation standard. This benchmark supplies all four.

What it measures that other benchmarks do not

This is not a generic detection benchmark with insects in it. It is built to separate failure modes that ordinary mAP conflates.

Detection and identification are scored separately. A detector here does two jobs — find the beetle, then name it. The benchmark reports detection recall, species accuracy given detection, and their product, so you can see which capability is failing. On curated imagery they fail for different reasons and on different species: on isolated specimens localization is close to solved (≈93% recall on the balanced in-distribution split, median IoU 0.98) while identification is not (≈60% of found specimens named correctly). A single mAP number hides that entirely.

Detection is not solved everywhere. The near-ceiling recall above is specific to curated photographs of individually mounted specimens. Recall falls with specimen density — on unseen-tribe species it is 99.8% for isolated specimens but 25.9% averaged over multi-specimen images, and 2.5% above ten specimens per frame — and collapses to 0.192–0.450 on field photographs. Dense-frame and field detection are open problems.

Fine-grained confusion is structured, not random. Species in the same genus are visually near-identical. Misidentifications land in the correct genus 12.6% of the time against 2.0% expected by chance (6.2× enrichment), and consistently misidentified species get absorbed into abundant look-alike labels. The full confusion matrix and the taxonomy for every species are included so this structure can be analysed rather than averaged away.

Presentation is a controlled variable. Specimens appear both as isolated individuals and on dense plates carrying tens or hundreds of beetles at once — the bulk output of ethanol-baited survey traps. Per-species imaging density is recorded, so you can ask whether a model fails on a taxon or on a photograph.

Taxonomic novelty is stratified. 110 species are held out entirely and banded by distance from the training set (within-genus, within-tribe, outside-tribe), which separates "can it find an unfamiliar beetle" from "can it name one".

Domain shift is measurable. A separate iNaturalist split contains field photographs of species that are in the training set, so photographic domain transfer is isolated from taxonomic transfer.

Annotation budget is a controlled axis. Three data-allocation regimes spanning seventeen budget conditions in total (six uniform-cap, six abundance-proportional, five target-acquisition) let you ask how a fixed number of expert-annotated images is best spent — breadth across species versus depth within them.

---

Splits

Split	Images	Scored annotations	Crowd regions	Specimens/image	Species	Purpose
train	6,096	53,040	0	8.70	65	trainable species
iid_test	620	650	4,367	8.09	65	held-out specimens, trained species
inat_test	74	80	8	1.19	8	field photographs, domain shift
semantic_ood	7,701	46,924	0	6.09	110	unseen species, banded by distance
All	14,491	100,694	4,375	7.25	175

Splits are disjoint at both image and annotation level: no source image identifier and no annotation record identifier is shared between any pair, and no individual beetle appears in more than one split.

iid_test and inat_test are balanced to 10 scored annotations per species, so no species dominates the mean and per-species AP is comparable across the label set.

Held-out species by taxonomic distance

Band	Species	Genera	Tribes	Images	Annotations	Effective species (Kish)
Within a genus present in training	59	23	13	4,227	23,791	11.4
Within a tribe, different genus	43	27	9	2,807	17,099	7.4
Outside any tribe present in training	8	5	4	667	6,034	2.6
All held-out	110	55	19	7,701	46,924	21.4

Bands are defined on taxonomy alone and are not matched on imaging composition. The outside-tribe band is the smallest and most densely imaged, and two species supply 84% of its annotations. Band-level comparisons must be disaggregated by density, and uncertainty on band-level claims should use a species-level cluster bootstrap rather than treating annotations as independent.

---

Crowd regions — read before evaluating

Because the test splits are balanced to 10 annotations per species, and some of those annotations were drawn from dense plates, other specimens on the same image are real but not scored. They are marked iscrowd=1 with their true category_id.

pycocotools handles this correctly with no changes: crowd regions are excluded from the recall denominator, and a detection landing on one is neither rewarded nor penalised.

Custom evaluation code must honour iscrowd. Treating crowd regions as ordinary targets inflates the iid_test ground truth from 650 to 5,017 and collapses recall — a detector that found every specimen and misplaced none could not exceed a precision of 0.130, and the penalty would fall entirely on species photographed in bulk. A crowd-aware reference evaluator is included at evaluation/evaluate.py, alongside evaluation/aggregate.py. Verify any evaluator before use:

grep -c "_n_real_gt" evaluation/evaluate.py   # expect 4 or more

The v1.0.0 evaluator is not crowd-aware and must not be used with v2 annotations.

train and semantic_ood are exhaustively annotated and contain no crowd regions.

---

Reference results

Five architectures at full training data (YOLOv8x, YOLOv10x, YOLO11x, YOLO12x, RT-DETR-X; mean of three seeds):

Metric	Range	What it tells you
iid_test AP@0.5	0.504 – 0.558	joint detect-and-name
iid_test detection recall	0.926 – 0.936	finding the beetle on curated imagery
Species accuracy given detection	0.581 – 0.626	naming it: the bottleneck
End-to-end species recall	0.538 – 0.586	what a deployed tool delivers
semantic_ood detection recall	0.797 – 0.810	finding an unfamiliar beetle
inat_test AP@0.5	13 – 30% of IID	domain transfer is poor
inat_test detection recall	0.192 – 0.450	field failure is localization, not naming
Novelty AUROC (confidence)	0.539 – 0.705	no usable abstention signal
FPR at 95% novelty recall	0.638 – 0.914	the cost of catching novel taxa

On curated imagery the gap between detection recall and species accuracy is the headline: models lose more than five times as much recall to misidentification (0.350–0.389) as to localization failure (0.064–0.074). Architecture choice barely moves this — per-species AP vectors correlate at ρ = 0.944–0.970 across architectures, so the ceiling is a property of the data and the task, not the backbone.

Under field photography the failure mode inverts: bounding-box geometry stays precise (median IoU 0.95) but mean confidence falls from 0.80 to 0.21 and the models miss the beetle entirely.

Four capabilities are unsolved and are what this benchmark is for measuring: discriminating congeners, detecting specimens in dense multi-specimen frames, recognising a species outside the label set, and detecting specimens under field photographic conditions.

---

Format

COCO detection JSON. Each image record carries source_image_uuid (links to the platform), width, height, and full licence metadata (license_name, license_url, license_verbatim, license_versioned). Each annotation carries source_record_id, category_id, absolute-pixel bbox, area, iscrowd, and depicts_valid_name_id.

from pycocotools.coco import COCO
coco = COCO("detection/annotations_coco/iid_test.json")
real = [a for a in coco.loadAnns(coco.getAnnIds()) if not a["iscrowd"]]
print(len(real))   # 650

Contents of this record

Path	What it is
detection/annotations_coco/	COCO JSON for each of the four splits
detection/images/	image files, redistributed unmodified
evaluation/evaluate.py	crowd-aware reference evaluator
evaluation/aggregate.py	seed and condition aggregation
image_licences.csv	per-image licence, photographer, holding institution
species_taxonomy.csv	full taxonomy and benchmark role for all 175 species
MANIFEST.csv	file list with SHA-256 digests
CHANGELOG.md	what changed in v2.0.1

---

Licensing

Images carry mixed Creative Commons licences and are redistributed unmodified. There is no single dataset-level licence. Images under licences that prohibit redistribution were excluded from this release.

Per-image licence, photographer and holding institution are in image_licences.csv, which is authoritative wherever it disagrees with this summary.

Some images come from sources that state no licence version. These are recorded as supplied with license_versioned: false and are not upgraded to 4.0. Users should satisfy the terms of the most restrictive plausible version. Unversioned licences concentrate in inat_test.

No no-derivatives (ND) images are present in any split. Images are redistributed unmodified as a Collection; ShareAlike obligations attach to adaptations rather than collections, so the SA-family images do not impose a licence on the dataset as a whole.

Annotations, splits, taxonomy and metadata are released under CC BY 4.0.

Users are responsible for verifying image-level attribution before redistribution or commercial use. The corpus cannot be treated as an unrestricted public-domain resource.

---

Known limitations

• Every annotated object is a beetle, so localization metrics measure whether an insect is found when present, not whether a detector can separate beetles from trap debris, wood fibres and non-target bycatch. Treat the reported recall as an upper bound for unsorted field collections.
• The outside-tribe band contains only eight species, so band-level novelty comparisons lack the power to resolve a subtle phylogenetic gradient in either direction.
• The iNaturalist split rests on 80 scored annotations across eight scolytine species and is indicative of the domain-shift mechanism rather than exhaustive.
• Isolated and multi-specimen presentations are intermixed across splits rather than fully crossed, so training and test density are not independently manipulable in this release.
• COCO scale-stratified metrics (AP small/medium/large) are uninformative here: all but 10 of the benchmark annotations fall in the "large" stratum. Characterize scale by specimens per image instead.

Citation

@dataset{marais_bark_ambrosia_benchmark_2026,
  title     = {Bark and Ambrosia Beetle Detection Benchmark},
  author    = {Marais, G. Christopher and Schuster, Layla A. and
               Johnson, Andrew J. and Kuo, Eric and Dias, Raquel and Hulcr, Jiri},
  year      = {2026},
  publisher = {Zenodo},
  version   = {2.0.1},
  doi       = {10.5281/zenodo.22695714}
}

See CHANGELOG.md for what changed in v2.0.1.