Occluded nuScenes

What It Is

Occluded nuScenes is a controlled multi-sensor occlusion extension of the nuScenes dataset.
The full paper title is "Occluded nuScenes: A Multi-Sensor Dataset for Evaluating Perception Robustness in Automated Driving."
It was released as a 2025 arXiv descriptor.
The dataset targets perception robustness under partial sensor failures and environmental occlusions.
It focuses on reproducible, parameterized degradations across camera, radar, and LiDAR.
It is a benchmark resource, not a new perception model.

Real datasets contain sensor noise, but they rarely provide controlled failure levels.
Occluded nuScenes starts from nuScenes scenes and adds parameterized occlusions.
Camera occlusions are released for both full and mini nuScenes versions.
Radar and LiDAR occlusions are generated with scripts so users can vary degradation parameters.
The benchmark enables repeatable comparisons between fusion models under the same occlusion conditions.
It isolates partial sensor failure from other dataset changes.
It is designed to support robustness analysis and safety-critical perception evaluation.

Inputs are nuScenes camera, radar, and LiDAR samples.
Camera inputs can be replaced with occluded image variants.
Radar and LiDAR inputs can be degraded through parameterized scripts.
Outputs are occluded multi-sensor samples compatible with nuScenes-style evaluation.
Downstream outputs depend on the tested perception model: boxes, tracks, segmentation, occupancy, or map elements.
The key benchmark output is performance as a function of sensor modality and occlusion severity.
The dataset assumes users already have access to nuScenes where required.

Camera modality includes four occlusion types.
Two camera occlusion types are adapted from public implementations.
Two camera occlusion types are newly designed by the authors.
Radar modality includes three parameterized degradation scripts.
LiDAR modality includes three parameterized degradation scripts.
Degradations can be applied independently to a modality or used to study fusion under multiple impaired sensors.
The protocol emphasizes controlled and reproducible adverse conditions.

The descriptor frames the resource as an evaluation dataset for automated driving perception robustness.
Typical usage is to evaluate a trained detector or fusion model on clean nuScenes and then on occluded variants.
Results should be reported by modality, occlusion type, and severity parameter.
The benchmark is especially useful for comparing camera-only, LiDAR-only, radar-only, and fused detectors.
It can be used to locate failure thresholds, such as percent occlusion where mAP or recall collapses.
It complements corruption benchmarks by emphasizing occlusion and partial sensor failure.
It should not be treated as a replacement for real contaminated-sensor datasets.

It provides controlled, repeatable sensor occlusion rather than relying on chance failures in logs.
It covers camera, radar, and LiDAR instead of camera-LiDAR only.
Full and mini camera releases support both lightweight and full-scale experiments.
Parameterized radar and LiDAR scripts make severity sweeps possible.
It is built on nuScenes, so many existing perception stacks can evaluate on it.
It helps separate true fusion resilience from overfitting to clean multi-sensor inputs.

Parameterized occlusions are not full physical simulations of dirt, water, ice, snow, or sensor electronics.
The resource inherits nuScenes object taxonomy and urban-road bias.
The arXiv descriptor is the primary source found here; release details should be verified before production use.
Occlusion alone does not test time skew, calibration drift, packet loss, or rolling shutter.
Multiple real failures can be correlated in ways that controlled scripts may not capture.
Radar and LiDAR occlusion parameters need mapping to physical sensor degradation before safety claims.
Results can be sensitive to whether models were trained with similar occlusions.

Occluded nuScenes is useful for designing airside sensor-occlusion test logic.
Camera occlusion maps to dirty lenses, rain streaks, de-icing spray, sun glare blocks, and service-vehicle blockage.
LiDAR and radar degradations map to blocked windows, spray, snow, and equipment shadowing.
Airside adaptation must add airport-specific sensor placements and occluders such as aircraft bodies and GSE.
The benchmark can help define graceful degradation thresholds for apron speed limits.
Production evidence still needs real airport sensor occlusion data and human-safety metrics.

Use clean nuScenes results as the baseline for every occlusion sweep.
Report both absolute performance and relative degradation.
Evaluate per-range and per-class behavior because near-field misses are safety critical.
Avoid training and testing on the same occlusion patterns when claiming generalization.
Log the exact script parameters used for radar and LiDAR occlusions.
Combine with MultiCorrupt and MSC-Bench to cover non-occlusion corruption families.