VisVerify-Bench / README.md
FaisaI's picture
Update README.md
f94c468 verified
|
Raw History Blame Contribute Delete
9.91 kB
metadata
license: mit
pretty_name: VisVerify-Bench
language:
  - en
size_categories:
  - 1K<n<10K
task_categories:
  - image-classification
  - object-detection
  - image-segmentation
tags:
  - annotation-verification
  - vlm-as-judge
  - decision-models
  - system-one
  - jev
  - benchmark
configs:
  - config_name: classification
    data_files:
      - split: dev
        path: data/classification/dev-*
      - split: test
        path: data/classification/test-*
  - config_name: detection
    data_files:
      - split: dev
        path: data/detection/dev-*
      - split: test
        path: data/detection/test-*
  - config_name: segmentation
    data_files:
      - split: dev
        path: data/segmentation/dev-*
      - split: test
        path: data/segmentation/test-*

VisVerify-Bench: judge images stamped accept or reject

VisVerify-Bench

A benchmark of vision annotation errors for VLM-as-judge and System One vision models.

Each item shows a judge an image with a proposed class, box or mask. Half the items are untouched ground truth and should be accepted. The other half were corrupted in a controlled way and should be rejected. Every wrong item records what is wrong (error_type) and how badly (iou_with_gt), so accuracy can be plotted against severity. Built to compare fast "System One" decision models (JEV, Liquid d1-omni) against VLM-as-judge baselines on the same items.

  • 3 tasks (classification, detection, segmentation) x 8 domains (everyday, street, aerial, medical, industrial, agriculture, documents/UI, wildlife/underwater)
  • 6,584 items: 5,988 test, 596 dev (separate photos, for prompt tuning only)
  • 50% correct / 50% corrupted in every task x domain cell; 16 error types, easy / medium / hard

Quick start

from datasets import load_dataset
import json

ds = load_dataset("FaisaI/VisVerify-Bench", "detection", split="test")
q = json.load(open("questions.json"))["detection"]          # question templates, from this repo

row = ds[0]
row["judge_image"]      # crop around the proposed box, 50% padding, box + class tag drawn
row["proposed_class"]   # "bus"
row["expected"]         # {"class_correct": True, "box_quality": 2, "annotation_correct": False}

Score answers against expected. Main metric: precision / recall / F1 on annotation_correct with reject as the positive class, plus the false-alarm rate on correct items.

Error types

Task error_type Corruption Severity
all class_swap_random / _similar / _cross_domain class replaced by a random / semantically close / other-domain class easy / hard / easy
det box_slightly_off / box_off / box_very_off box moved to IoU 0.75-0.85 / 0.50-0.75 / 0.20-0.50 hard / medium / easy
det box_wrong_object / box_background box on another object / on empty background medium / easy
seg mask_eroded / mask_dilated morphology to IoU 0.60-0.85 medium / hard
seg mask_missing_part / mask_leak / mask_shifted 20-40% chunk removed / grown into background / translated to IoU 0.40-0.70 medium / medium / easy
seg mask_wrong_instance mask of another instance of the same class hard
det, seg class_and_box / class_and_mask class swap and a geometry error easy

20% of correct boxes and masks carry a tiny jitter (IoU >= 0.92, tag tp_jittered) so "pixel-perfect" cannot be used as a shortcut. Quality scores map from IoU: 5 (>= 0.90), 4 (0.75), 3 (0.60), 2 (0.40), 1 (below).

Fields

id, task, split, domain, source_dataset, source_id, license · image (original; the crop for classification), judge_image (default rendering) · proposed_class, class_description, gt_class · proposed_bbox, gt_bbox as [x, y, w, h] pixels · proposed_mask, gt_mask as COCO RLE · is_correct, error_type, severity, iou_with_gt · choice_options, choice_answer (4-way question, classification) · expected · tags · notes

Questions live once in questions.json. Images are at most 1024 px on the long side; raw geometry is kept so you can re-render (overlay_full, crop_tight, crop_context, side_by_side in the code repo). At most 3 items per photo, no class above 10% of a domain.

Counts

classification

domain none class_swap_cross_domain class_swap_random class_swap_similar total
aerial 165 81 43 41 330
agriculture 165 79 43 43 330
documents_ui 165 55 55 55 330
everyday 165 55 55 55 330
industrial 165 55 55 55 330
medical 165 55 55 55 330
street 165 55 55 55 330
wildlife_underwater 165 55 55 55 330
total 1320 490 416 414 2640
domain dev test
aerial 30 300
agriculture 30 300
documents_ui 30 300
everyday 30 300
industrial 30 300
medical 30 300
street 30 300
wildlife_underwater 30 300

Severity: easy=906, hard=414, none=1320

detection

domain none box_background box_off box_slightly_off box_very_off box_wrong_object class_and_box class_swap_cross_domain class_swap_random class_swap_similar total
aerial 165 18 18 18 18 18 24 17 17 17 330
agriculture 165 19 18 18 18 18 24 17 17 16 330
documents_ui 165 18 18 18 18 18 24 17 17 17 330
everyday 165 18 18 18 18 18 24 17 17 17 330
industrial 165 19 18 18 18 17 24 18 17 16 330
medical 165 18 18 18 18 18 24 17 17 17 330
street 165 18 18 18 18 18 24 17 17 17 330
wildlife_underwater 165 18 19 18 18 17 24 18 17 16 330
total 1320 146 145 144 144 142 192 138 136 133 2640
domain dev test
aerial 30 300
agriculture 30 300
documents_ui 30 300
everyday 30 300
industrial 30 300
medical 30 300
street 30 300
wildlife_underwater 30 300

Severity: easy=756, hard=277, medium=287, none=1320

segmentation

domain none class_and_mask class_swap_cross_domain class_swap_random class_swap_similar mask_dilated mask_eroded mask_leak mask_missing_part mask_shifted mask_wrong_instance total
everyday 165 24 17 17 16 16 15 15 15 15 15 330
medical 165 0 28 0 0 28 28 27 27 27 0 330
street 157 23 16 16 16 15 15 15 14 14 13 314
wildlife_underwater 165 24 17 17 16 16 15 15 15 15 15 330
total 652 71 78 50 48 75 73 72 71 71 43 1304
domain dev test
everyday 30 300
medical 30 300
street 26 288
wildlife_underwater 30 300

Severity: easy=270, hard=166, medium=216, none=652

Sources and licenses

Source Where License Redistributable Tasks Notes
COCO 2017 val images.cocodataset.org Annotations CC BY 4.0; images carry individual Flickr CC licenses Yes, with per-image filter cls, det, seg Keep only images whose COCO license id is 4 (CC BY 2.0), 5 (CC BY-SA 2.0), 7 (no known copyright) or 8 (US Gov). Record the per-image license in license.
LVIS v1 val dl.fbaipublicfiles.com/LVIS Annotations CC BY 4.0; same COCO images Yes, same filter cls, det, seg 1,203 classes, good for class_swap_similar.
Open Images V7 validation storage.googleapis.com/openimages Annotations CC BY 4.0; images CC BY 2.0 Yes cls, det, seg Masks come from the V5 mask release (validation-masks-*.zip + CSV). Google does not guarantee each image's license; attribute per image.
RF100-VL HF probicheaux/rf100-vl Apache 2.0 (stated in roboflow/rf100-vl repo) Yes cls, det 100 subsets across domains. Boxes only: checked 2026-10-30, the *-seg subsets carry no polygons.
PlantDoc HF agyaatcoder/PlantDoc CC BY 4.0 Yes cls, det Original: github.com/pratikkayal/PlantDoc-Object-Detection-Dataset.
ISIC 2016 Part 1 HF MedOtter/ISIC2016 CC0 1.0 Yes cls, det, seg 1,279 dermoscopy images with expert lesion masks. Single class, so classification distractors come from cross-domain swaps.
DocLayNet v1.1 HF ds4sd/DocLayNet-v1.1 CDLA-Permissive 1.0 Yes cls, det Some HF mirrors tag license: other; the card body and the IBM page say CDLA-P 1.0.
PubLayNet HF creative-graphic-design/PubLayNet CDLA-Permissive 1.0 Yes cls, det

No non-commercial source is used. COCO and LVIS photos keep their individual Flickr licenses (only CC BY, CC BY-SA, public domain and US Government photos were used) and each row records its own. Segmentation covers 4 domains; aerial, industrial and agriculture had no permissively licensed mask source, and documents is detection-only.

Citation

@misc{visverifybench2026,
  title  = {VisVerify-Bench: Vision Annotation Errors for VLM Judges and System One Models},
  author = {Faisal},
  year   = {2026},
  url    = {https://huggingface.co/datasets/FaisaI/VisVerify-Bench}
}

Sources: COCO (Lin et al. 2014), LVIS (Gupta et al. 2019), Open Images (Kuznetsova et al. 2020), Roboflow100-VL (Robicheaux et al. 2025), PlantDoc (Singh et al. 2020), ISIC 2016 (Gutman et al. 2016), DocLayNet (Pfitzmann et al. 2022), PubLayNet (Zhong et al. 2019).