Chris1095/clean-kitti-2d-object-detection

Cleaned KITTI Dataset

Overview

This repository provides a cleaned version of the KITTI 2D Object Detection Benchmark. The dataset was curated as part of the research project:

Analyzing Training-Free Corruption Detection for Object Detection Datasets

The goal of the cleaning process was to identify potential annotation inconsistencies using a training-free feature-space based corruption detection approach.

The original KITTI annotation format and folder structure are preserved.

Cleaning Methodology

Potential annotation errors were identified using feature-space similarity analysis based on pretrained visual embedding models. All detected samples were manually reviewed and corrected when appropriate.

Detected corruption types include:

• Mislabel: incorrect semantic class assignments
• Badly located: imprecise bounding boxes
• Other annotation inconsistencies

The method is particularly effective at identifying semantic inconsistencies. However, it has limited sensitivity towards small or moderate positional deviations of bounding boxes. Therefore, remaining annotation errors are expected.

Dataset Usage

This dataset is intended as a research artifact for:

• Benchmarking annotation quality analysis methods
• Studying real-world annotation inconsistencies
• Reproducing the experiments presented in the associated publication

It should not be considered a perfectly clean ground-truth version of KITTI.

Citation

If you use this dataset, please cite both:

1. The original KITTI publication

   Geiger, A., Lenz, P., and Urtasun, R.
   Are we ready for autonomous driving? The KITTI Vision Benchmark Suite. CVPR 2012.

2. The publication introducing this cleaned version: Link

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

Acknowledgements

All images and original annotations originate from the KITTI Vision Benchmark Suite. This repository only contains the modifications required for the cleaning and analysis process.