Monocular Depth Estimation Using 2D-to-3D Feature Lifting

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing monocular depth estimation systems inaccurately represent depth due to undetected occlusions and lack of effective spatial relationship detection.

Innovation Solution

A hybrid neural network architecture combining a 2D CNN for feature extraction and a 3D CNN for depth estimation, utilizing a feature lifting module to transform 2D features into 3D inputs, enabling accurate detection of occlusions and spatial relationships.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If a single image is used for depth estimation, then the system complexity is reduced, but the depth estimation accuracy deteriorates due to undetected occlusions and insufficient spatial relationship detection

Engineering Contradiction:
Improvesystem complexityVSAvoiddepth estimation accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent transforms 2D image features into 3D feature representations by introducing a temporal dimension through video sequences. By processing multiple frames over time, the system creates 3D feature volumes that enable detection of occlusions and spatial relationships without requiring complex multi-camera setups, thus resolving the contradiction between system simplicity and depth estimation accuracy

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Solution Approach 2:

The system performs preliminary feature extraction and lifting operations on video sequences before final depth estimation. By pre-processing the data to create 3D feature representations and detect occlusions in advance, the system improves depth accuracy while maintaining computational efficiency during the actual depth estimation process

Inventive Principle:
Principle #10Preliminary action

2Productivity

If traditional 2D CNN features are used directly for depth estimation, then the processing is simpler, but the detection of occlusions and spatial relationships is insufficient

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidocclusion detection capability
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent introduces a feature lifting module as an intermediary between 2D CNN extraction and depth estimation. This module transforms 2D features into 3D feature representations, enabling reliable occlusion and spatial relationship detection while maintaining processing efficiency. The intermediary layer bridges the gap between simple 2D processing and complex 3D analysis without requiring full 3D CNN processing

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

By lifting 2D features to 3D feature space through the feature lifting module, the system gains the ability to detect occlusions and spatial relationships reliably. This dimensional transformation allows the system to process video sequences efficiently while achieving robust occlusion detection, resolving the contradiction between processing efficiency and detection reliability

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Data Source

PatentEP4081951B1Systems and methods for depth estimation
Publication Date: 2026.04.01 GOOGLE LLC
  • EP4081951B1 patent drawingFigure 1
  • EP4081951B1 patent drawingFigure 2
  • EP4081951B1 patent drawingFigure 3

AI summary

A system includes a computing device. The computing device is configured to perform a set of functions. The set of functions includes receiving an image, wherein the image comprises a two-dimensional array of data. The set of functions includes extracting, by a two-dimensional neural network, a plurality of two-dimensional features from the two- dimensional array of data. The set of functions includes generating a linear combination of the plurality of two-dimensional features to form a single three-dimensional input feature. The set of functions includes extracting, by a three-dimensional neural network, a plurality of three-dimensional features from the single three-dimensional input feature. The set of functions includes determining a two-dimensional depth map. The two-dimensional depth map contains depth information corresponding to the plurality of three-dimensional features.