RobustMMD: towards robust multimodal 3D detection via efficient multilevel adapter and spatial denoiser

Accurate 3D object detection is critical for autonomous driving. While multi-modal fusion techniques leverage complementary information from multiple sensors to overcome the limitations of single-modal approaches, existing methods are vulnerable to input noise and sensor calibration errors, leading to significant performance degradation. This is largely because current models lack dedicated mechanisms to handle noisy inputs. To address this limitation, a Multi-Level Adapter and a Multi-modal Semantic-Guided Spatial Denoiser are proposed for robust multi-modal 3D object detection under noisy conditions. Integrating these modules into existing detectors significantly enhances robustness to noisy inputs with minimal computational overhead. Moreover, the proposed approach can be seamlessly incorporated into mainstream multi-modal detection frameworks. On the NuScenes-C benchmark, our method improves the baseline methods’ mAP by 18.6%~54.0% under strong sunlight noise.