Chris1095/cleaned_voc2012_annotations

Cleaned VOC2012 Dataset

Overview

This repository provides a curated version of the PASCAL VOC2012 object detection dataset.

The dataset combines two independent cleaning stages:

1. Hasty.ai quality-controlled annotations.
2. Additional corrections based on training-free feature-space corruption detection as presented in:

Analyzing Training-Free Corruption Detection for Object Detection Datasets

The objective of this repository is not to provide a definitive error-free benchmark, but to provide a reproducible research artifact for studying annotation quality and corruption detection in object detection datasets.

---

Dataset Structure

VOC2012/
├── Annotations/
│   └── *.xml
│
├── hasty_cleaned_annotations/
│   ├── combined/
│   ├── train/
│   ├── validation/
│   ├── test/
│   └── Clean_PASCAL_COCO_Format.json
│
├── training_free_cleaned_annotations/
│   ├── annotations/
│   ├── image_examples/
│   │   ├── badly_located/
│   │   ├── mislabel/
│   │   └── others/
│   │
│   ├── correction_report.json
│   └── Cleaned_PASCAL_COCO_Format.json

The Annotations directory contains the original PASCAL VOC2012 annotations.

The hasty_cleaned_annotations directory contains the quality-controlled annotations provided by Hasty.ai.

The training_free_cleaned_annotations directory contains the corrections generated during this work, including correction reports and visual examples of modified annotations.

---

Cleaning Methodology

Potential annotation inconsistencies were identified using a training-free feature-space-based corruption detection pipeline.

The pipeline operates on individual object instances by extracting bounding-box crops and analyzing their similarity within a feature space generated by pretrained visual embedding models.

Detected annotations were manually inspected and categorized into:

• Mislabel: Incorrect semantic class assignment.
• Badly located: Bounding boxes that do not accurately enclose the object.
• Other: Remaining annotation inconsistencies.

The approach is effective at identifying semantic inconsistencies but remains less sensitive to positional errors. Therefore, despite the applied cleaning stages, remaining annotation errors may still exist.

---

Dataset Statistics

The Hasty.ai cleaned version contains:

• 17,119 images
• 43,294 annotated objects

During our additional inspection, 63 remaining annotation errors were identified and removed:

• 24 mislabels
• 15 badly located bounding boxes
• 24 other inconsistencies

---

Intended Use

This dataset is intended for:

• Research on dataset auditing.
• Evaluation of annotation corruption detection methods.
• Controlled experiments involving synthetic annotation noise.
• Reproducibility of the experiments presented in the associated publication.

It should not be treated as a guaranteed ground-truth dataset.

---

Citation

If you use this dataset, please cite:

1. The original PASCAL VOC publication.

2. The Hasty.ai PASCAL cleaning work.

3. The publication introducing the training-free cleaning procedure. Link

   Sieberichs, C., Geerkens, S., Waschulzik, T., Viswanathan, R., and Braun, A. Analyzing Training-Free Corruption Detection for Object Detection Datasets. DataCV 2026

---

Acknowledgements

This repository only provides additional annotation files, documentation, and corrections generated during the cleaning process. The images can be found at the original VOC (https://www.robots.ox.ac.uk/~vgg/projects/pascal/VOC/) and are not provided within this repository.