Monocular Depth Estimation Using Mean and Confidence Outputs

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing depth estimation technologies require stereo images and complex image matching processes, limiting their efficiency and accuracy, especially in complex scenes.

Innovation Solution

A neural network apparatus that estimates depth using a monocular image, outputting both a mean depth value and a confidence level, allowing for improved accuracy and efficiency through probabilistic modeling and selective training based on confidence levels.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If stereo images and image matching processes are used for depth estimation, then depth information can be obtained based on geometric relationships, but the device complexity and computational requirements increase significantly

Engineering Contradiction:
Improvedepth estimation accuracyVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent extracts and utilizes geometric cues directly from monocular images, removing the need for stereo image pairs and complex matching processes. The neural network is trained to identify and leverage geometric relationships (such as vanishing points, horizon lines, and object proportions) present within a single image to estimate depth, thereby simplifying the system while maintaining depth estimation capability

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent replaces the mechanical/optical system of stereo cameras and matching algorithms with a computational neural network system. Instead of using geometric matching between two images, a single neural network processes one image to directly output depth estimates, substituting complex computational geometry operations with learned patterns from training data

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Measurement precision

If stereo images are used for depth estimation, then depth information can be obtained, but the processing time and computational resources increase

Engineering Contradiction:
Improvedepth information accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The neural network is pre-trained on large datasets containing monocular images with corresponding depth information. During training, the network learns to directly map image features to depth values without requiring runtime matching operations. This preliminary training phase enables fast inference on unseen images, as the depth estimation becomes a direct function evaluation rather than an iterative matching process

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent removes the time-consuming image matching step entirely by extracting depth information directly from monocular image features through the trained neural network. The system processes single images through learned feature representations to produce depth maps, eliminating the sequential operations of stereo pairing, feature matching, and disparity calculation that characterize traditional methods

Inventive Principle:
Principle #2Taking out (Extraction)

3Productivity

If monocular depth estimation is used, then processing efficiency improves, but depth estimation accuracy deteriorates in complex scenes

Engineering Contradiction:
Improveprocessing efficiencyVSAvoiddepth estimation accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The neural network undergoes extensive preliminary training on diverse and complex scene data, learning to handle challenging cases such as occlusions, varying lighting conditions, and complex geometries. This pre-training equips the model with robust feature representations and depth prediction capabilities that generalize well to complex scenes during deployment, maintaining accuracy while preserving the efficiency benefits of monocular processing

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent employs data augmentation techniques during training that systematically vary image parameters such as lighting conditions, camera angles, scene complexity, and object configurations. By training on extensively parameter-diverse data, the network learns invariant features and robust depth estimation patterns that maintain accuracy across complex scenes. The loss function and training parameters are also optimized to specifically address challenging depth estimation cases

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3945497B1Method and apparatus with image depth estimation
Publication Date: 2026.02.04 SAMSUNG ELECTRONICS CO LTD
  • EP3945497B1 patent drawingFigure 1
  • EP3945497B1 patent drawingFigure 2
  • EP3945497B1 patent drawingFigure 3

AI summary

A method and apparatus with image depth estimation are provided. The method includes obtaining a first statistical value associated with a depth for each of plural pixels included in an input image based on a first channel of output data obtained by applying the input image to a neural network, obtaining a second statistical value associated with a depth for each of the plural pixels in the input image based on a second channel of the output data, and estimating depth information of each of the plural pixels in the input image based on the first statistical value and the second statistical value. The neural network may be trained based on a probability distribution for a depth of each pixel in an image based on a first statistical value and a second statistical value that are obtained corresponding to an image with predetermined depth information in the training data.