Depth Estimation Loss Function for Uncertain Regions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional depth estimation techniques using loss functions like L1 and BerHu loss place greater emphasis on regions with large estimation errors, which often correspond to uncertain depths, leading to inaccurate depth estimation in images.
Innovation Solution
A depth estimation apparatus that uses a loss function where the error increase is smaller for pixels with errors larger than a threshold, using a radical root function for larger errors and a linear function for smaller errors, to reduce the emphasis on uncertain regions and improve estimation accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional loss functions (L1, BerHu) are used that place greater emphasis on regions with large estimation errors, then the loss function can theoretically address difficult estimation cases, but the depth estimation accuracy deteriorates because large errors often correspond to uncertain depths
Solution Approach 1:
The patent applies local quality by making the loss function's treatment of errors spatially variable through the uncertainty map. Instead of uniformly penalizing all large errors, the system selectively applies loss only to regions where uncertainty is below a threshold, thereby adapting the loss function's strictness to local confidence levels across the depth map.
Solution Approach 2:
The patent changes the parameter of the loss function by introducing an uncertainty-based threshold that dynamically controls which error regions contribute to the loss. This transforms the static loss function into a dynamic one that adapts its behavior based on local uncertainty characteristics, effectively filtering out unreliable error regions.
2Measurement precision
If multiple images from different viewpoints are used to estimate depth, then depth information can be obtained through triangulation, but the device complexity and operational requirements increase
Solution Approach 1:
The patent uses a neural network to learn a mapping from single images to depth maps, effectively copying the depth estimation capability from trained examples rather than requiring physical multiple-view cameras. The network is trained on paired image-depth data and then applies this learned knowledge to estimate depth from single images.
Solution Approach 2:
The patent replaces the mechanical/optical system of multiple cameras and triangulation geometry with a computational neural network system. Instead of using physical baseline separation and geometric computation, the system uses learned patterns from training data to directly predict depth, substituting mechanical depth sensing with intelligent algorithms.
3Device complexity
If a single image is used for depth estimation, then the apparatus complexity is reduced, but the depth estimation accuracy deteriorates compared to multi-view methods
Solution Approach 1:
The patent applies preliminary action by pre-training the neural network on extensive paired image-depth data before deployment. This offline training phase prepares the network with learned depth cues and patterns, enabling it to perform accurate depth estimation during actual operation without requiring additional hardware or complex processing at runtime.
Solution Approach 2:
The neural network copies depth estimation knowledge from training examples to new input images. By learning the mapping between image features and depth values during training, the network can replicate accurate depth estimation performance on unseen images, effectively copying the expertise gained during training.
Data Source
AI summary
An acquiring unit of a depth estimation apparatus acquires an input image. In addition, a depth map generating unit inputs the input image acquired by the acquiring unit into a depth estimator for generating, from an image, a depth map in which a depth of a space that appears on the image is imparted to each pixel of the image, and generates an estimated depth map that represents a depth map corresponding to the input image. The depth estimator is a model having been learned in advance so as to reduce, with respect to each error between a depth of the estimated depth map and a depth of a correct-answer depth map that presents the depth map of a correct answer, a value of a loss function set such that a degree of increase of a loss value with respect to a pixel at which the error is larger than a threshold is smaller than a degree of increase of a loss value with respect to a pixel at which the error is equal to or smaller than the threshold.