PhD-DETR: Physics-guided depth-aware transformer for monocular 3D object detection under adverse weather conditions

Robust monocular 3D object detection under adverse weather conditions remains a significant challenge due to monocular depth ambiguity and weather-induced visual degradation. In this paper, we propose PhD-DETR (Physics-guided Depth-aware Transformer), a physics-guided reliability-constrained framework for robust monocular 3D object detection. Instead of using physical priors as fixed replacements for learned visual features, PhD-DETR formulates environment and frequency priors as adaptive soft reliability cues within a depth-aware Transformer detector. Specifically, PhD-DETR introduces a Prior-knowledge Gated Fusion Module to adaptively integrate image features and aligned physical priors, a Physics-guided Adaptive Attention Mechanism to modulate deformable attention logits with query-specific prior responses, and a Physics-guided Attention Regularization loss to encourage object-related attention distributions to be consistent with physically meaningful cues. Experiments on KITTI, KITTI-C and V2X-Radar-V demonstrate that PhD-DETR improves robustness under both controlled synthetic corruptions and real-world adverse sensing conditions. Extensive ablation studies further show that the proposed physical priors provide complementary reliability cues rather than merely acting as generic spatial masks or redundant feature enhancements.