A stereo matching algorithm based on the improved PSMNet.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) have advanced stereo matching, but existing models suffer from high parameter counts and poor performance in ill-posed regions.
• Challenges include reflections, repeated textures, and fine structures, leading to inaccurate disparity estimation in traditional and some deep learning methods.

Purpose of the Study:

• To develop a lightweight and efficient stereo matching network that addresses the limitations of existing CNN-based approaches.
• To improve disparity estimation accuracy in challenging areas with weak or repeated textures and reflections.

Main Methods:

• A lightweight Pyramid Stereo Matching Network (PSMNet) was developed, incorporating ResNeXt for feature extraction and Atrous Spatial Pyramid Pooling (ASPP) for multiscale features.
• A feature fusion module was designed to construct a matching cost volume, followed by a 3D CNN with an encoder-decoder structure for regularization.
• Disparity maps were generated via regression, and the model was evaluated on Scene Flow, KITTI 2012, and KITTI 2015 datasets.

Main Results:

• The proposed network demonstrated comparable prediction accuracy to PSMNet on benchmark datasets.
• A significant reduction in running time was achieved compared to the original PSMNet, highlighting its efficiency.
• Improved performance was observed in ill-conditioned areas, such as those with repeated or weak textures.

Conclusions:

• The lightweight stereo matching network effectively balances accuracy and speed, outperforming PSMNet in processing time.
• The model successfully handles challenging scenarios in stereo matching, offering a more robust solution for real-world applications.
• This research contributes a more efficient and accurate deep learning approach to stereo matching problems.