Stereo Matching Using Multi-Level Cost Volume and Multi-Scale Feature Constancy.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Cost volumes are crucial for accurate stereo matching in Convolutional Neural Networks (CNNs).
• Existing methods often do not fully leverage the information contained within cost volumes.
• Developing robust stereo matching models that generalize across diverse datasets remains a challenge.

Purpose of the Study:

• To propose an end-to-end trainable CNN that maximizes the utility of cost volumes for stereo matching.
• To enhance the efficiency and accuracy of disparity refinement.
• To develop a robust model capable of performing well on multiple datasets with varying characteristics.

Main Methods:

• An end-to-end CNN architecture comprising shared feature extraction, initial disparity estimation, and disparity refinement sub-modules.
• Cost volume calculation at multiple levels using shared features, integrated into both estimation and refinement stages.
• Introduction of multi-scale feature constancy for efficient disparity refinement and a two-stage finetuning scheme for cross-dataset robustness.

Main Results:

• The proposed method achieves state-of-the-art performance on widely recognized benchmarks: Middlebury 2014, KITTI 2015, ETH3D 2017, and SceneFlow.
• The model demonstrated superior accuracy and robustness across datasets with different characteristics.
• The method secured first place in the Stereo task at the Robust Vision Challenge 2018.

Conclusions:

• The proposed CNN effectively utilizes cost volumes for high-accuracy stereo matching.
• The network architecture and training strategy enable robust performance across diverse datasets.
• This work advances the state-of-the-art in deep learning-based stereo matching.