Monocular Depth Image Estimation With Instance Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for obtaining depth images, such as binocular vision and time of flight, often result in inaccurate depth estimation, which can compromise the safety of autonomous vehicles by failing to accurately monitor pedestrians and obstacles.
Innovation Solution
A method utilizing a deep learning network to estimate depth and pose from monocular images, combined with instance segmentation and transformation matrix averaging, to enhance depth estimation accuracy by segmenting moving and non-moving objects and refining transformation matrices, ultimately converting depth maps to point cloud images for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If binocular vision or time of flight methods are used to obtain depth images, then depth information can be captured, but the depth estimation accuracy is insufficient
Solution Approach 1:
The patent replaces traditional mechanical/optical depth estimation methods (binocular vision, time of flight) with a deep learning-based computational method. The deep learning network processes monocular images to directly predict depth maps, achieving superior accuracy without requiring additional hardware cameras or complex optical systems.
Solution Approach 2:
The patent transforms the depth estimation problem from physical measurement parameters (optical flow, time of flight) to data-driven parameters (deep learning network outputs). By training the network on labeled depth data, the system learns to map image pixels to depth values, fundamentally changing how depth is measured from physical optics to computational modeling.
2Measurement precision
If deep learning network is used to estimate depth from monocular images, then depth estimation accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent performs preliminary action by pre-training the deep learning network on a large dataset of labeled images before actual depth estimation. This pre-training phase captures complex depth relationships in the network weights, allowing the model to make accurate predictions during inference without requiring complex real-time computations or additional processing steps.
Solution Approach 2:
The patent creates a computational model (deep learning network) that copies and learns from the complex relationships between images and depth information. Instead of physically measuring depth through complex optical systems, the network copies the patterns and relationships learned from training data to predict depth, simplifying the physical system while maintaining accuracy.
Data Source
AI summary
A method for obtaining depth images implemented in an electronic device includes obtaining a first image and a second image; obtaining a predicted depth map of the first image, and calculating a first error value of the predicted depth map; determining a first transformation matrix between the first image and the second image; obtaining an instance segmentation image and obtaining a first mask image and a second mask image; obtaining a target transformation matrix; converting the predicted depth map into a first point cloud image, converting the first point cloud image into a second point cloud image, and converting the second point cloud image into a third image; calculating a second error value between the second image and the third image; obtaining a target deep learning network model; and inputting at least one image into the target deep learning network model, and obtaining at least one depth image.