Monocular Depth Estimation With Mixed Supervision for Absolute Depth

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Monocular depth estimation techniques require frequent fine-tuning and additional sensors, leading to increased costs and complexity, and existing methods struggle to provide accurate absolute depth predictions without ground truth data.

Innovation Solution

A mixed supervision training process is employed using both labeled and unlabeled datasets to train a monocular depth estimation network, combining fully-supervised and self-supervised training methods to enhance depth estimation accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If monocular depth estimation techniques use additional sensors (LiDAR, radar, stereo cameras) to improve depth estimation accuracy, then measurement precision is improved, but device complexity and cost increase

Engineering Contradiction:

Improvedepth estimation accuracyVSAvoidsensor system complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent extracts and removes the dependency on additional sensors (LiDAR, radar, stereo cameras) from the depth estimation system. By using only monocular images from a single camera, the system eliminates complex sensor hardware while maintaining depth estimation capability through learned features from training data.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent uses copying by training the monocular depth estimation network on source data that includes ground truth depth information from multiple sensors, then applies this learned knowledge to target data from single-camera inputs. The network copies depth estimation patterns from labeled source images to unlabeled target images without requiring the target to have additional sensors.

Inventive Principle:

Principle #26Copying

2Adaptability or versatility

If monocular depth estimation techniques require frequent fine-tuning with new scene data, then adaptability is improved, but loss of time and productivity decrease

Engineering Contradiction:

Improvescene adaptation capabilityVSAvoidfine-tuning time

Core Design Contradiction:

Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent performs preliminary action by pre-training the depth estimation network on diverse source data encompassing multiple scenes and conditions. This pre-training equips the network with generalized depth estimation knowledge that can be quickly adapted to new target scenes through the mixed supervision process, reducing the need for extensive fine-tuning when encountering new environments.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent changes parameters by using a mixed supervision training approach that combines labeled source data with unlabeled target data. This parameter change in training strategy allows the network to adapt to new scenes by leveraging both supervised signals from source data and self-supervised signals from target data, achieving fast adaptation without time-consuming manual fine-tuning.

Inventive Principle:

Principle #35Parameter changes

3Ease of operation

If monocular depth estimation uses only unlabeled target data for training, then ease of operation is improved, but measurement precision deteriorates due to inability to provide absolute depth predictions

Engineering Contradiction:

Improvetraining data requirementVSAvoidabsolute depth prediction accuracy

Core Design Contradiction:

Ease of operationVSMeasurement precision

Solution Approach 1:

The patent merges labeled source data with unlabeled target data in a mixed supervision training framework. The network simultaneously learns from supervised signals (source images with ground truth depth maps) and self-supervised signals (target images without ground truth), combining the advantages of both approaches to achieve accurate absolute depth predictions while maintaining ease of operation with minimal labeled data requirements.

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The patent uses the labeled source data as an intermediary to transfer depth estimation knowledge to the unlabeled target data. The mixed supervision process acts as a mediator that bridges the gap between supervised and unsupervised learning, allowing the network to leverage ground truth information from source data to improve depth prediction accuracy on target data without requiring direct ground truth annotations for target scenes.

Inventive Principle:

Principle #24Intermediary (Mediator)

Data Source

PatentUS20260073537A1Transferring salient depth properties from labeled data to unlabeled datasets for monocular depth estimation

Publication Date: 2026.03.12 SAMSUNG ELECTRONICS CO LTD

AI summary

A method and apparatus for training a monocular depth estimation (MDE) network, including: obtaining a source dataset including a first source image and a first ground truth depth map corresponding to the first source image; obtaining a target dataset comprising a first target image and a second target image; generating an estimated first source depth map corresponding to the first source image using the MDE network; generating an estimated target depth map corresponding to the first target image using the MDE network; generating an estimated relative pose based on the first target image and the second target image using a pose network; and training the MDE network and the pose network by performing mixed supervision training, wherein the performing the mixed supervision training includes performing fully-supervised training based on the estimated first source depth map and the first ground truth depth map, and performing self-supervised training based on the estimated target depth map and the first estimated relative pose