The KITTI Vision Benchmark Suite

We propose a lightweight RGB-D road segmentation
scheme from three complementary dimensions: cross-
modal feature rectification, high-level context
modeling, and data-level robustness enhancement.
First, a Difference-driven Road-aware cross-modal
Feature Rectification Module (DR-FRM) is designed,
which explicitly captures the consensus and
disagreement between RGB and depth features. It
adaptively suppresses noisy modal features while
enhancing reliable ones across both channel and
spatial dimensions, thereby mitigating error
propagation caused by coarse-grained fusion.
Second, a Depthwise Large-kernel Pyramidal
Aggregation module (DLPA) is proposed, which
explicitly expands the effective receptive field
of high-level features with minimal additional
computational overhead, improving the modeling of
contextual information for long-range lane
markings, curves, and complex road topologies.
Third, the cross-dataset instance-level RGB-D
augmentation strategy introduced in Evi-RoadSeg is
use

lr=1e-4
weight_decay=1e-4
optimizer=AdamW
epochs=400