An efficient and accurate multi-level cascaded recurrent network for stereo matching.

Area of Science:

• Computer Vision
• Deep Learning
• Robotics

Background:

• Transformer-based convolutional neural networks achieve high accuracy in stereo matching for disparity estimation.
• Current state-of-the-art methods suffer from significant model inference time, limiting practical applications.

Purpose of the Study:

• To address the accuracy-efficiency trade-off in stereo matching.
• To propose an efficient and accurate multi-level cascaded recurrent network for disparity estimation.

Main Methods:

• Designed a multi-level network for coarse-to-fine recurrent iterative updates of difference values.
• Introduced slow-fast multi-stage superposition inference structures for diverse scene data.
• Incorporated adaptive and lightweight group correlation layers to enhance speed and reduce rectification errors.

Main Results:

• The proposed LMCR-Stereo network achieves competitive disparity estimation accuracy.
• Demonstrated significant improvements in model inference speed: 46.0% on SceneFlow and 50.4% on Middlebury.
• The method effectively balances accuracy and efficiency compared to existing state-of-the-art approaches.

Conclusions:

• LMCR-Stereo offers a viable solution for real-time stereo matching applications.
• The network architecture effectively enhances disparity estimation accuracy and inference speed.
• The proposed method contributes to advancing efficient computer vision and robotics systems.