Beyond Isolated Corruptions: Benchmarking 3D Robust Perception
across Compositional Adversity Space

Chunmian Lin^*,†, Haobing Pang^*, Lei Yang, Xu Han, Daxin Tian^†

^* Equal contribution · ^† Corresponding authors

Paper

Code

Data

Abstract

Performance under nominal sensing conditions alone offers an incomplete picture of perception robustness. Existing corruption benchmarks largely focus on isolated perturbations or a limited set of mixtures, leaving the interaction between heterogeneous degradations insufficiently characterized. We introduce ComboBench, a compositional robustness benchmark built on a configurable generation pipeline, ComboGen. The pipeline organizes over 30 atomic operators into five mechanism-grounded taxonomies and synthesizes physically compatible compositions with controlled operator order and severity. We further propose CARE, a Corruption-Aware Routing Expert framework that combines degradation-aware expert selection, corruption-responsive feature adaptation, and reliability-guided multimodal fusion. Experiments on nuScenes-Combo and KITTI-Combo diagnose the vulnerability of 18 clean-trained detectors and assess robustness under category-balanced mixtures and free-form compositions. Ablations and visualizations examine component contributions, routing behavior, and generalization to unseen combinations.

Highlights

Atomic operators: 30+
Corruption taxonomies: 5
Severity levels: 5+
Datasets: 2
Detectors diagnosed: 18

Configurable compositional benchmark. ComboGen assembles over 30 atomic operators into physically compatible compositions, with control over selection, order, and severity.
Corruption-aware expert adaptation. CARE coordinates degradation-aware routing, expert modulation, and cross-modal reliability calibration.
Systematic robustness evaluation. KITTI-Combo and nuScenes-Combo support individual-corruption diagnosis, compositional assessment, and analysis of unseen combinations.

Corruption Samples

Compositional severity

Fog + Echo Depletion + Occlusion · Increasing severity

Severity 1

Severity 3

Severity 5+

CARE Architecture

Experimental Results

Robustness Evaluation

We evaluate 18 clean-trained detectors on nuScenes-Combo and KITTI-Combo without additional training or test-time tuning. Detection and robustness metrics are averaged over the applicable corruption types and severity levels, revealing distinct response profiles across camera-based, LiDAR-only, and multimodal detectors.

Figure 5: six radar plots of detection and robustness metrics for clean-trained detectors on KITTI-Combo and nuScenes-Combo

Robustness Enhancement

CARE is integrated into representative camera-based, LiDAR-only, and multimodal detectors to evaluate robustness under compound sensing degradations. The following results compare CARE-equipped models with their corresponding baselines and representative alternatives under the joint Fog + Echo Depletion + Occlusion composition on both validation benchmarks.

Original Table 7: nuScenes-Combo free-form corruption results comparing base and CARE-enhanced detectors

Original Table 8: KITTI-Combo free-form corruption results comparing base and CARE-enhanced detectors

Qualitative Results

nuScenes-combo

Fog-Echo Depletion-Occlusion

BibTeX

@misc{combobench,
  title  = {Beyond Isolated Corruptions: Benchmarking 3D Robust
            Perception across Compositional Adversity Space},
  author = {Lin, Chunmian and Pang, Haobing and Yang, Lei
            and Han, Xu and Tian, Daxin}
}

Beyond Isolated Corruptions · Citation