When rain and blur degrade the camera, the fusion layer learns to trust it less: 1.41 mAP gain on nuScenes rain
Synopsis
Camera–LiDAR bird's-eye-view fusion can mistake corrupted visual texture for useful evidence when the camera stream degrades. DARFusion recasts fusion as modality trust estimation: learned reliability cues calibrate both feature streams, and calibrated features, task-driven contexts and a compact degradation descriptor derived from camera BEV features decide how much camera evidence to use. It holds 68.4 mAP and 71.5 NDS on the clean nuScenes validation set and beats the reproduced BEVFusion baseline by 1.41 and 1.31 mAP on real rain and night subsets.
Interpretation
Camera–LiDAR bird's-eye-view fusion now carries a separate judgment: the camera is still producing features, but whether its evidence should be trusted is estimated rather than read off response strength. Earlier fusion modules concatenate or convolve features by content, so a strong activation counts as useful semantics; under glare, noise, blur or low light the same strong response can come from corrupted texture or sensor artifacts. On the nuScenes validation set, DARFusion lifts the reproduced BEVFusion baseline from 67.8 mAP and 71.0 NDS to 68.4 mAP and 71.5 NDS, both means over three random seeds.
The fusion weight is set jointly by the calibrated feature pair, task-driven auxiliary contexts, and a degradation descriptor read from the camera BEV feature. That descriptor maps three response statistics of a projected response map, global activation mean, edge response and local high-frequency residual, through learnable affine transforms into 0 to 1, requiring no degradation labels or image-quality supervision. On the nuScenes real rain subset of 1088 samples and night subset of 602 samples, DARFusion gains 1.41 and 1.31 mAP over BEVFusion, both above the 0.59 mAP gain on the full validation set.
As degradation grows, the fusion layer gives the camera branch a smaller effective share, a trend seen for synthetic night, motion blur and Gaussian noise alike. On 128 held-out nuScenes samples, the dedicated descriptor correlates positively with severity rank while the magnitude-based camera contribution ratio correlates negatively. Spearman correlations with severity are 0.70, 0.65 and 0.55 for the descriptors under synthetic night, motion blur and Gaussian noise, and -0.62, -0.57 and -0.42 for the camera contribution ratio, with 95% scene-bootstrap intervals excluding zero.
Perspective
Evaluation covers the clean nuScenes validation set, real rain and night subsets, three controlled camera corruptions (synthetic night, motion blur, Gaussian noise), and nine degraded KITTI settings; in the controlled settings LiDAR and calibration parameters stay unchanged, and models are trained only on clean data.
The degradation descriptors are task-driven conditioning cues rather than calibrated physical-quality measurements, and the mechanism analysis reports statistical associations, not causal claims; at the strongest Gaussian noise the two methods are nearly tied, and synthetic corruptions are proxies for real imaging degradation.