Adaptive Kernel Convolutional Stereo Matching Recurrent Network.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Current advanced stereo matching relies on iterative structures using Gated Recurrent Units (GRUs).
• Existing GRU-based methods often lack non-local geometric and contextual information in cost volumes.
• This limitation hinders performance, especially in complex scenarios.

Purpose of the Study:

• To propose a novel deep recurrent network architecture for stereo matching.
• To enhance cost volume representation by incorporating adaptive kernel convolutions and attention mechanisms.
• To improve the accuracy and generalization ability of binocular stereo matching.

Main Methods:

• Developed a GRU iteration-based adaptive kernel convolution deep recurrent network (AKC-Stereo).
• Introduced a kernel convolution-based adaptive multi-scale pyramid pooling (KAP) module to capture spatial correlations.
• Incorporated a matching attention (MAR) module to refine cost volumes before iterative updates.

Main Results:

• The AKC-Stereo network demonstrated superior performance compared to the basic network.
• Achieved an End-Point Error (EPE) of 0.45 on the Sceneflow dataset, a 0.02 improvement.
• Outperformed the base network by 5.6% on the D1-all metric for the KITTI 2015 dataset.

Conclusions:

• The proposed AKC-Stereo network effectively enhances stereo matching by integrating adaptive kernel convolutions and attention mechanisms.
• The KAP and MAR modules significantly improve pixel-level representation and network generalization.
• This approach offers a substantial advancement in binocular stereo matching accuracy.