Cross-modal semantic scene completion for complex urban traffic scenes with reachability geometric attention

In complex urban roads, mutual occlusions among multiple entities are pervasive. Approaches that rely solely on visible-appearance segmentation or perform completion only in voxel space struggle to recover pixel-level semantics and shapes for occluded regions. This paper presents AGAC, a cross-modal, temporally aware semantic scene completion network augmented with reachability-aware geometric attention. We first construct an anisotropic reachability potential field in a unified reference frame and differentiably project both the potential values and the steepest-descent directions onto the image plane, yielding per-pixel reachability scalars, preferred propagation directions, and reachability masks. We then introduce reachability-aware geometric attention in the pixel domain, embedding potential differences, directional consistency, and occlusion risk into both the attention neighborhood and its scoring, so that long-range completion propagates only along geometrically feasible, low-cost, and directionally consistent conduits. The decoder performs hierarchical occlusion parsing with anisotropic refinement, diffusing along the principal direction while suppressing boundary leakage along the normal, thereby achieving amodal filling with continuous contours and topologically consistent shapes. Finally, a geometry-reprojection-gated dual-window memory with uncertainty weighting delivers temporally stable cross-frame fusion and robust reliability quantification on hard cases. System-level evaluations on KITTI-360-APS, BDD100K-APS, and BlendPASS show that AGAC achieves leading performance on APQ, APC, AAP, and occluded-instance AP_occ, improving overall metrics by approximately 3-4% over strong baselines and AP_occ by about 5-7%. Gains are more pronounced under heavy occlusion, large scale, and long range, while quality in visible regions remains stably unimpaired.