Deep Stereo Matching With Hysteresis Attention and Supervised Cost Volume Construction

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Stereo matching is crucial for depth sensing and autonomous driving.
• Existing end-to-end networks use a pipeline of feature extraction, cost volume construction, aggregation, and regression.

Purpose of the Study:

• To propose a novel deep neural network architecture for stereo matching.
• To improve the feature extraction and cost volume construction stages of the stereo matching pipeline.

Main Methods:

• Developed a multiple-block attention module inspired by hysteresis comparators.
• Constructed the cost volume in a supervised manner.
• Employed a data-driven approach to balance feature map informativeness and compactness.

Main Results:

• The proposed method was evaluated on SceneFlow, KITTI 2012, and KITTI 2015 datasets.
• Achieved superior performance compared to previous methods on these benchmark datasets.

Conclusions:

• The novel attention mechanism and supervised cost volume construction effectively enhance stereo matching.
• The proposed architecture offers improved disparity prediction accuracy for vision tasks.