Sub-pixel Depth Estimation via Class Likelihood Parabola
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional depth estimation techniques using stereo images struggle with accurate estimation of intermediate depth classes, as they typically only output one class closest to the true depth, failing to estimate sub-pixel units and requiring numerous classes for high accuracy, which complicates neural network scaling and regularization.
Innovation Solution
A learning device that calculates a likelihood for multiple estimated classes using an estimation model, learns a loss function with weights applied to a function having a predetermined shape, such as a parabola curve or equiangular line, to estimate classes in sub-pixel units, enabling higher accuracy by determining the extreme value position corresponding to the correct data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional depth estimation techniques output only one class closest to the true depth, then the estimation process is simple, but the measurement precision of depth estimation deteriorates
Solution Approach 1:
The patent changes the output parameters of the estimation model from discrete class labels to continuous probability distributions over multiple classes. By outputting likelihoods for multiple depth classes simultaneously and using these in a parabola curve function, the system achieves sub-pixel depth estimation precision without proportionally increasing model complexity
Solution Approach 2:
The patent transitions from one-dimensional discrete class output to a two-dimensional continuous probability space by applying a parabola curve function to multiple class likelihoods. This dimensional transformation enables precise depth estimation at sub-pixel levels while maintaining computational efficiency
2Measurement precision
If numerous depth classes are used to achieve high accuracy, then the measurement precision of depth estimation is improved, but the device complexity and difficulty of regularization increase
Solution Approach 1:
Instead of increasing the number of discrete depth classes, the patent changes the parameter representation from discrete class counts to continuous probability likelihoods. This allows the model to achieve high precision by computing weighted combinations of likelihoods across existing classes, avoiding the complexity of scaling to numerous classes
3Measurement precision
If conventional techniques estimate only integer pixel units, then the estimation model is simpler, but the measurement precision of depth estimation deteriorates
Solution Approach 1:
The patent introduces a continuous dimension by applying a parabola curve function to discrete class likelihoods. This enables the model to estimate depths at sub-pixel precision levels (between integer pixel units) by finding the extreme value position in the continuous function, achieving higher precision without proportionally increasing computational resources
Data Source
AI summary
A learning device includes one or more processors. The processors calculate a likelihood of belonging to a plurality of estimated classes, of learning data, by using an estimation model for estimating to which of the estimated classes input data belongs. The processors calculate a weight of a loss function to be used in learning the estimation model such that, when a likelihood of a first class that is closer to correct data than other estimated classes among the estimated classes and likelihoods of a second class and a third class that are adjacent to the first class are applied to a function having a predetermined shape, a position that has an extreme value of the function corresponds to the correct data. The processors learn the estimation model by using the loss function.