Monocular Depth Map Generation Using Encoder-Decoder DCNN
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional 3D depth estimation techniques are cumbersome, power-intensive, and costly, making them unsuitable for small robots and drones, which require compact, low-cost, and low-power solutions for accurate three-dimensional recognition.
Innovation Solution
A deep convolutional neural network (DCNN) optimized for an encoder-decoder architecture is used to generate depth maps from monocular images, extracting features and calculating displacements of mismatched pixels to produce high-resolution depth maps with reduced computational requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If active type depth estimation using laser scanner or structured light is used, then measurement precision is improved, but device complexity and cost increase
Solution Approach 1:
The patent replaces active mechanical depth sensing systems (laser scanners, structured light sensors) with a passive computational approach using a single camera and deep learning algorithms. The DCNN model processes monocular images to estimate depth, substituting complex hardware with software-based processing that achieves comparable accuracy without requiring active illumination or specialized sensors.
2Measurement precision
If stereo matching is used for passive depth estimation, then measurement precision is improved, but use of energy and device complexity increase
Solution Approach 1:
The patent extracts depth information from a single monocular image by training the DCNN to infer depth cues from monocular visual data alone, eliminating the need for stereo matching computations. This extraction approach focuses computational resources on processing one image stream while achieving depth estimation through learned features rather than exhaustive pixel correspondence searches.
3Measurement precision
If conventional depth sensors are used, then measurement precision is improved, but weight and power consumption increase
Solution Approach 1:
The patent creates a computational copy of depth sensing capability through a deep learning model that processes standard monocular images. Instead of using physical depth sensors (LiDAR, structured light cameras), the system uses a DCNN that learns to predict depth maps from regular 2D images, effectively copying the depth sensing function through software rather than specialized hardware.
4Productivity
If DCNN with encoder-decoder architecture is used, then productivity is improved, but device complexity increases
Solution Approach 1:
The patent segments the DCNN into distinct encoder and decoder modules with specific functions. The encoder extracts hierarchical features from input images through multiple convolutional layers, while the decoder reconstructs depth maps from these features. This segmentation allows for optimized computation at each stage and enables efficient implementation on resource-constrained devices while maintaining high processing speed.
Data Source
AI summary
Disclosed is an apparatus for generating a depth map using a monocular image. The apparatus includes: a deep convolution neural network (DCNN) optimized based on an encoder and decoder architecture. The encoder extracts one or more features from the monocular image according to the number of provided feature layers, and the decoder calculates displacements of mismatched pixels from the features extracted from different feature layers, and generates the depth map for the monocular image.