Disparity Estimation via Weakly Supervised Cost Propagation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing binocular disparity estimation methods face challenges in accurately estimating dense disparity, particularly in regions with specular reflection, low light, transparency, and no texture, due to high parameter adjustment requirements, data dependency, and poor generalization across scenarios.
Innovation Solution
A binocular disparity estimation method that combines traditional feature matching with weakly supervised deep learning, using a three-dimensional convolutional network to optimize initial cost diagrams and convert them into probability-based dense disparity maps, thereby addressing issues of false matching and data label dependency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional binocular disparity estimation methods are used, then sparse disparity estimation is accurate, but dense disparity estimation has severe defects in regions of specular reflection, low light, transparency and no texture
Solution Approach 1:
The patent combines traditional binocular disparity estimation methods with deep learning methods to create a hybrid approach. The traditional method provides accurate sparse disparity estimation through geometric constraints, while the deep learning component (using 3D convolutional networks) fills in the dense disparity information, particularly improving performance in challenging regions like specular reflection, low light, transparency, and no texture areas.
Solution Approach 2:
The patent creates a composite estimation system that integrates two different methodological 'materials': traditional geometric constraint-based methods and data-driven deep learning methods. This composite approach leverages the strengths of each method while compensating for their individual weaknesses, achieving both sparse and dense disparity estimation accuracy.
2Measurement precision
If deep learning based disparity estimation methods are used, then feature expression ability is stronger and disparity diagram is more accurate, but data dependency is strong and generalization ability is poor
Solution Approach 1:
The patent performs preliminary action by using the traditional method to obtain accurate sparse disparity estimates first. These sparse estimates are then used as constraints or guidance for the deep learning model, allowing the model to learn from limited labeled data while being guided by geometrically accurate references, thus improving generalization ability.
Solution Approach 2:
The patent changes the parameter of data supervision from fully supervised to weakly supervised. By using sparse disparity labels from traditional methods as supervision signals rather than requiring dense pixel-level labels, the model achieves good generalization ability while maintaining high accuracy on diverse datasets.
3Ease of operation
If methods combining deep learning with traditional method are used, then interpretability is improved, but full advantages of both methods are not utilized and accuracy advantage over end-to-end learning is not shown
Solution Approach 1:
The patent introduces a new dimension in the combination approach by using 3D convolutional networks to process the cost volume in an additional dimensional space. This allows the model to maintain interpretability through the traditional cost aggregation framework while achieving end-to-end learning accuracy through the 3D convolutional feature learning, effectively utilizing advantages of both traditional and deep learning methods.
Data Source
AI summary
The present invention discloses a disparity estimation method for weakly supervised trusted cost propagation, which utilizes a deep learning method to optimize the initial cost obtained by the traditional method. By combining and making full use of respective advantages, the problems of false matching and difficult matching of untextured regions in the traditional method are solved, and the method for weakly supervised trusted cost propagation avoids the problem of data label dependency of the deep learning method.