Under adverse weather conditions, camera-based depth estimation often becomes unreliable due to reduced contrast and severe visual degradation. Radar measurements remain stable under these conditions, but they are extremely sparse and provide limited ...
Under adverse weather conditions, camera-based depth estimation often becomes unreliable due to reduced contrast and severe visual degradation. Radar measurements remain stable under these conditions, but they are extremely sparse and provide limited structural cues. Because the reliability of the two sensors changes differently depending on the environment, existing fusion methods often fail to maintain consistent depth scales, especially at long ranges.
To address this issue, this work proposes Adaptive Dual-Confidence Fusion Network(ADCNet), a camera-radar fusion model that explicitly estimates and calibrates the confidence of each sensor before fusion. ADCNet first predicts camera and radar confidences separately to capture reliability variations under rain or snow. The predicted confidences are then calibrated using temperature scaling to mitigate over-confident predictions. Finally, the calibrated confidences are normalized with a softmax function and used to adaptively fuse the two sensor features.
We evaluate ADCNet on the Boreas dataset, which includes a wide range of adverse-weather scenes. Compared with AFNet and existing camera-radar fusion methods(Li et al., CaFNet), ADCNet consistently reduces MAE, RMSE, iMAE, iRMSE, and AbsRel. Qualitative results further demonstrate that ADCNet produces more stable and visually consistent depth maps in heavy rain and snow than existing methods.
Overall, the results indicate that modeling sensor reliability and adaptively fusion camera and radar improves robustness under challenging weather, and that even sparse radar signals can stabilize depth estimation when integrated appropriately.