Gradient-Domain Disparity Networks for Stereo Boundary Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for determining distances from stereoscopic images, particularly near object boundaries, are not always accurate, leading to inconsistencies in computer vision and robotics applications.
Innovation Solution
A deep learning-based disparity network utilizing end-to-end training with gradient-domain smoothness priors and occlusion reasoning, combined with pixel-adaptive convolutions, to enhance distance estimation accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional stereoscopic image processing methods are used, then the processing speed is fast, but the distance determination accuracy deteriorates near object boundaries and edges
Solution Approach 1:
The patent introduces gradient-domain smoothness priors as an intermediary mechanism to mediate between the left and right stereoscopic images. By computing gradients in the gradient domain and enforcing smoothness constraints, the system resolves ambiguities at object boundaries without requiring complex post-processing, thus improving distance determination accuracy while maintaining reasonable system complexity
Solution Approach 2:
The patent transforms the problem from direct pixel-space matching to gradient-domain processing. By operating in the gradient domain rather than the original image domain, the system captures fine details and edge information more effectively, improving accuracy at object boundaries while the neural network handles the dimensional transformation automatically
2Measurement precision
If neural networks with extensive ground truth data are used, then distance determination accuracy improves, but the data requirement and training complexity increase
Solution Approach 1:
The patent implements self-service through the gradient-domain smoothness prior module, which automatically enforces smoothness constraints during the disparity estimation process without requiring external ground truth data for supervision. The system uses its own gradient computations to guide the learning process, reducing dependency on extensive annotated training data while maintaining high accuracy
Solution Approach 2:
The patent changes the fundamental parameters of the learning problem by operating in the gradient domain rather than the intensity domain. This parameter transformation allows the neural network to learn more effective features with fewer training examples, as gradient information provides stronger geometric constraints that reduce the need for large amounts of ground truth data
3Measurement precision
If pixel-adaptive convolutions are applied, then fine detail capture improves, but the computational complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the image processing into distinct functional modules: gradient computation, smoothness prior enforcement, and disparity estimation. The pixel-adaptive convolutions are localized to specific layers within the neural network where they are most needed, rather than applying them uniformly across the entire processing pipeline, thus reducing overall computational energy while maintaining fine detail capture
Data Source
AI summary
Apparatuses, systems, and techniques are presented to determine distance for one or more objects. In at least one embodiment, a disparity network is trained to determine distance data from input stereoscopic images using a loss function that includes at least one of a gradient loss term and an occlusion loss term.