The KITTI Vision Benchmark Suite

Stereo matching algorithms struggle to adapt to
highly reflective and textureless surfaces
Propose a lightweight, learnable context-aware group
correlation computation module: CAGWC

The model was implemented using the PyTorch deep
learning framework, and all experiments were
conducted on NVIDIA RTX 4090 GPUs. For training
the stereo matching network, the AdamW optimizer
was employed, with gradients of learnable
parameters clipped to the range of [-1, 1]. When
training on the SceneFlow dataset, the total
number of training steps was set to 200k, using a
one-cycle learning rate scheduling policy with an
initial learning rate of $2\times10^{-4}$. For
training on the KITTI dataset, the model was
initialized with weights pre-trained on SceneFlow
and then fine-tuned on the KITTI dataset for 60k
steps, also using a one-cycle learning rate
scheduler with an initial learning rate of
$1\times10^{-4}$. For all experiments, the batch
size was set to 12. Input images were randomly
cropped to a size of 256 $\times$ 768 pixels, and
data augmentation techniques such as asymmetric
color augmentation and spatial augmentation were
applied. During training, 22 update iterations
were