The KITTI Vision Benchmark Suite

This work proposes DCR-RoadSeg, a lightweight RGB-
D road segmentation network that jointly addresses
cross-modal feature fusion, high-level context
modeling, and data-level robustness. A Difference-
driven Road-aware Feature Rectification Module
(DR-FRM) exploits the consensus and disagreement
between RGB and depth to adaptively recalibrate
features in both channel and spatial dimensions,
while a Depthwise Large-kernel Pyramidal
Aggregation (DLPA) module efficiently enlarges the
receptive field of high-level features for complex
road geometries. In addition, a geometrically
consistent instance-level RGB-D augmentation
strategy based on cross-dataset copy–paste
enriches occlusion- and illumination-challenging
cases and enhances robustness on small-scale
datasets. Extensive experiments on benchmark road
datasets demonstrate that DCR-RoadSeg achieves
superior accuracy and robustness compared with
existing methods, while maintaining real-time
inference.

lr=1e-4
optimizer=AdamW
epochs=600