pith. sign in

arxiv: 2605.09774 · v1 · submitted 2026-05-10 · 💻 cs.CV

DRIVE-C: A Controlled Corruption Dataset for Autonomous Driving

Shiva Aher This is my paper

Pith reviewed 2026-05-12 02:04 UTC · model grok-4.3

classification 💻 cs.CV
keywords autonomous drivingcorruption datasetperception robustnesscamera degradationsensor health monitoringout-of-distribution detectionsynthetic degradationsvideo dataset
0
0 comments X

The pith

DRIVE-C supplies pixel-aligned pairs of clean and synthetically corrupted driving videos to benchmark how autonomous perception systems respond to known camera degradations.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces DRIVE-C as a structured testbed built from real forward-facing driving videos across diverse conditions. Clean clips are anonymized and then transformed using physics-inspired synthetic degradations into 600 corrupted versions spanning twelve types at five severity levels. This produces matched clean-degraded pairs with full metadata and sensor health annotations, allowing direct measurement of perception changes under controlled inputs rather than unpredictable real failures. A sympathetic reader would see value in using these reproducible cases to isolate and quantify the effects of specific degradations on downstream tasks like object detection or lane keeping.

Core claim

DRIVE-C is a dataset of ten clean real-world driving clips and six hundred corrupted counterparts that applies twelve camera degradation types at five severity levels to create pixel-aligned pairs, complete with per-clip metadata and Global Sensor Health Index annotations, so that perception reliability can be evaluated under fully reproducible controlled corruption conditions.

What carries the argument

The DRIVE-C construction process that starts with anonymized real driving videos and applies physics-inspired synthetic degradations to generate pixel-matched clean and corrupted versions together with reproducible parameters and sensor health labels.

If this is right

  • Perception models can be benchmarked on identical inputs with and without each specific degradation to isolate failure modes.
  • Systems can be trained or tuned to remain accurate when particular camera degradations appear at known severity levels.
  • Out-of-distribution detection methods gain a controlled reference set for labeling inputs as degraded versus nominal.
  • Sensor health monitoring algorithms can use the Global Sensor Health Index annotations to validate their predictions against ground-truth corruption parameters.
  • Reproducible test clips enable consistent comparison of different autonomous driving pipelines under the same degradation conditions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the synthetic degradations prove representative, the dataset could serve as a standard reference for regulatory or industry robustness certification of ADAS components.
  • The paired structure makes it straightforward to measure the exact performance cost of each degradation type, which could guide hardware choices such as camera placement or redundancy.
  • Extending the same alignment approach to additional sensor modalities like lidar or radar would create multi-modal controlled corruption benchmarks.

Load-bearing premise

The chosen synthetic degradations accurately capture the visual effects and interactions of actual camera failures that occur during autonomous driving.

What would settle it

Collect real-world instances of camera failures in driving footage and measure whether the resulting visual artifacts or drops in perception model accuracy differ substantially from the patterns produced by the twelve synthetic types in DRIVE-C.

Figures

Figures reproduced from arXiv: 2605.09774 by Shiva Aher.

Figure 1
Figure 1. Figure 1: Teaser of DRIVE-C. A clean forward-facing driving clip is paired with controlled corrupted variants across [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of DRIVE-C. Real forward-facing driving videos are anonymized and converted into clean clips, then [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ground-truth GSHI as a function of degradation [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Predicted versus ground-truth GSHI across all 610 [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
read the original abstract

DRIVE-C is a controlled corruption dataset designed to evaluate visual perception robustness in autonomous driving systems. It is built from real-world forward-facing driving videos collected across daytime, nighttime, urban, rural, freeway, and parking environments. Clean clips are anonymized via localized face and license plate blurring, then transformed with physics-inspired synthetic degradations. The dataset contains 10 clean clips and 600 corrupted clips spanning 12 camera degradation types across five severity levels, with per-clip metadata and Global Sensor Health Index (GSHI) annotations. DRIVE-C supports robustness benchmarking, degradation-aware modeling, uncertainty estimation, out-of-distribution (OOD) detection, and sensor health monitoring for Advanced Driver Assistance Systems (ADAS). By providing pixel-aligned clean and degraded video clips with fully reproducible corruption parameters, DRIVE-C offers a structured testbed for studying perception reliability under controlled camera degradation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 3 minor

Summary. The manuscript introduces DRIVE-C, a controlled corruption dataset for autonomous driving perception. It consists of 10 clean real-world forward-facing driving video clips collected across daytime/nighttime, urban/rural, freeway, and parking environments. These are anonymized via localized face and license plate blurring, then transformed using 12 physics-inspired synthetic camera degradations at five severity levels to produce 600 corrupted clips. The dataset provides pixel-aligned clean/degraded pairs, per-clip metadata, and Global Sensor Health Index (GSHI) annotations. It is intended to support robustness benchmarking, degradation-aware modeling, uncertainty estimation, OOD detection, and sensor health monitoring for ADAS by enabling reproducible experiments under controlled degradations.

Significance. If the dataset is released with verified pixel alignment and fully specified reproducible corruption parameters, DRIVE-C would provide a useful structured testbed for isolating the effects of specific camera degradations on perception systems. Controlled synthetic corruptions on real driving footage allow systematic evaluation that is hard to achieve with purely real-world data, facilitating research on robustness, uncertainty, and sensor monitoring. The addition of GSHI annotations extends utility toward practical ADAS health assessment. This aligns with needs in safety-critical autonomous driving where understanding degradation impacts is essential.

major comments (2)
  1. The exact formula or aggregation method for computing the Global Sensor Health Index (GSHI) from corruption parameters and metadata is not specified in the dataset description or annotations section. This is load-bearing for the claim that DRIVE-C supports sensor health monitoring, as users cannot interpret or extend the index without its definition.
  2. The manuscript states 600 corrupted clips spanning 12 degradation types and five severities but lacks a table or breakdown (e.g., in the dataset statistics section) showing distribution across the six environment types. Without this, it is unclear whether the claimed diversity is achieved in a balanced way that supports the testbed utility for cross-environment robustness studies.
minor comments (3)
  1. Add example frames or short video snippets illustrating each of the 12 degradation types at low, medium, and high severity to help readers visualize the controlled corruptions.
  2. Clarify in the pipeline description whether the anonymization blurring is applied before or after degradation synthesis, and confirm that it does not compromise pixel-level alignment between clean and corrupted versions.
  3. Include a statement on data and code release plans, including exact parameter files, to substantiate the 'fully reproducible corruption parameters' claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive recommendation of minor revision and for the constructive comments, which help improve the clarity and utility of the DRIVE-C dataset description. We address each major comment below.

read point-by-point responses

  1. Referee: The exact formula or aggregation method for computing the Global Sensor Health Index (GSHI) from corruption parameters and metadata is not specified in the dataset description or annotations section. This is load-bearing for the claim that DRIVE-C supports sensor health monitoring, as users cannot interpret or extend the index without its definition.

    Authors: We agree that the explicit formula for GSHI must be provided in the manuscript to support the sensor health monitoring use case. The current version describes the presence of GSHI annotations but does not detail the aggregation. In the revised manuscript we will insert a dedicated subsection (under Dataset Annotations) that states the precise formula, including how corruption parameters, severity levels, and per-clip metadata are combined into the index. This will allow users to interpret, reproduce, and extend GSHI. revision: yes

  2. Referee: The manuscript states 600 corrupted clips spanning 12 degradation types and five severities but lacks a table or breakdown (e.g., in the dataset statistics section) showing distribution across the six environment types. Without this, it is unclear whether the claimed diversity is achieved in a balanced way that supports the testbed utility for cross-environment robustness studies.

    Authors: We acknowledge that an explicit breakdown is needed to substantiate the claimed environmental diversity. The manuscript currently asserts coverage across daytime/nighttime, urban/rural, freeway, and parking but does not tabulate the counts. In the revised version we will add a table in the Dataset Statistics section that reports the number of clips per environment type, further stratified by degradation type and severity level. This will confirm balanced representation and strengthen the argument for cross-environment robustness evaluation. revision: yes

Circularity Check

0 steps flagged

No significant circularity; dataset construction is self-contained

full rationale

The paper is a dataset release describing collection of real driving videos, anonymization via blurring, and application of physics-inspired synthetic degradations to produce pixel-aligned clean/degraded pairs with metadata. No equations, fitted parameters, predictions, or derivation chains exist. The central claim (provision of a controlled, reproducible testbed) follows directly from the pipeline description without reduction to inputs by construction or self-citation. No load-bearing steps require external verification beyond the stated reproducibility of the transforms.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

The central contribution is a dataset rather than a derivation, so the ledger contains no free parameters, no mathematical axioms, and no invented physical entities. The only new construct is the GSHI annotation, which is defined within the paper and has no independent external validation shown.

invented entities (1)
  • Global Sensor Health Index (GSHI) no independent evidence
    purpose: Per-clip scalar annotation indicating overall sensor health under degradation
    Introduced as a new annotation layer; no external validation or prior reference is provided in the abstract.

pith-pipeline@v0.9.0 · 5431 in / 1199 out tokens · 36436 ms · 2026-05-12T02:04:31.243116+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages · 1 internal anchor

  1. [1]

    Vi- sion meets robotics: The kitti dataset,

    A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, “Vi- sion meets robotics: The kitti dataset,”International Journal of Robotics Research, vol. 32, no. 11, pp. 1231–1237, 2013

    work page 2013

  2. [2]

    nuscenes: A multimodal dataset for au- tonomous driving,

    H. Caesar, V. Bankiti, A. H. Lang, S. Vora, V. E. Liong, Q. Xu, A. Krishnan, Y. Pan, G. Baldan, and O. Beijbom, “nuscenes: A multimodal dataset for au- tonomous driving,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, 2020, pp. 11 621–11 631

    work page 2020

  3. [3]

    Scalability in perception for au- tonomous driving: Waymo open dataset,

    P. Sunet al., “Scalability in perception for au- tonomous driving: Waymo open dataset,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 2446–2454

    work page 2020

  4. [4]

    Benchmarking neu- ral network robustness to common corruptions and perturbations,

    D. Hendrycks and T. Dietterich, “Benchmarking neu- ral network robustness to common corruptions and perturbations,” inProceedings of the International Conference on Learning Representations, 2019

    work page 2019

  5. [5]

    Descrip- tor: Simon fraser university electric vehicle parking dataset (sfu-evp),

    S. Makonin, I. Ziyat, R. Sampson, S. An, F. Popowich, P. Palmer, and D. Agosti, “Descrip- tor: Simon fraser university electric vehicle parking dataset (sfu-evp),”IEEE Data Descriptions, vol. 1, pp. 13–21, 2024

    work page 2024

  6. [6]

    Drive-c: A controlled corruption dataset for autonomous driving,

    S. Aher, “Drive-c: A controlled corruption dataset for autonomous driving,” https://doi.org/10.5281/ zenodo.19656444, 2026, zenodo dataset

    work page 2026

  7. [7]

    Dji osmo action 4 product specifications,

    DJI, “Dji osmo action 4 product specifications,” https://www.dji.com/osmo-action-4/specs, 2024

    work page 2024

  8. [8]

    Ultralyt- ics yolo,

    G. Jocher, A. Chaurasia, and J. Qiu, “Ultralyt- ics yolo,” https://github.com/ultralytics/ultralytics, 2024

    work page 2024

  9. [9]

    The opencv library,

    G. Bradski, “The opencv library,”Dr. Dobb’s Jour- nal of Software Tools, 2000

    work page 2000

  10. [10]

    Pytorch: An imperative style, high- performance deep learning library,

    A. Paszkeet al., “Pytorch: An imperative style, high- performance deep learning library,” inAdvances in Neural Information Processing Systems, 2019

    work page 2019

  11. [11]

    Imagehash,

    J. Buchner, “Imagehash,” https://github.com/ JohannesBuchner/imagehash, 2024

    work page 2024

  12. [12]

    Safety-Critical Camera Reliability Monitoring for ADAS via Degradation-Aware Uncertainty Pattern Analysis

    S. Aher, “Safety-critical camera reliability moni- toring for adas via degradation-aware uncertainty pattern analysis,” https://arxiv.org/abs/2605.05439, 2026, arXiv preprint. 9

    work page internal anchor Pith review Pith/arXiv arXiv 2026