The KITTI Vision Benchmark Suite

This method jointly performs semantic segmentation
and stereo disparity estimation. Geometry-guided
Semantic Modulation injects disparity-derived
geometric cues into the semantic branch, while
Semantic-guided Disparity Refinement uses semantic
cues to selectively refine disparity estimates. A
Boundary Geometry Coupling loss provides class-
pair-aware supervision at semantic boundaries
during training without increasing inference cost.
The full model contains 8.22M parameters and runs
at 67.37 FPS on an NVIDIA A100 GPU.

max_disp=192; GSM window K=5; alpha, beta, and
gamma are learned;
lambda_seg=lambda_disp=lambda_bgc=1; lambda_b=1;
lambda_g=0.5; tau=2% of the training-set semantic
boundary samples.