Depth Estimation Using Confidence-Weighted Ground Truth Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for estimating depth information from images face challenges in generating complete depth information due to incomplete data collection, necessitating the use of depth completion technologies, which can result in inaccurate ground truth depth maps.
Innovation Solution
A method and apparatus that utilize a confidence map for a ground truth depth map to improve depth information estimation by learning a depth information estimation model, incorporating the confidence map into the loss function to enhance accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Area of stationary object
If depth completion technology is used to generate complete depth information from incomplete Lidar data, then the coverage of depth map is improved, but the accuracy of ground truth depth map deteriorates
Solution Approach 1:
The patent applies local quality by creating a confidence map that assigns different confidence levels to different regions of the depth map. Completed depth values have lower confidence values while actual measured values have higher confidence values, allowing the system to treat different regions differently based on their reliability
Solution Approach 2:
The patent changes the parameter representation by introducing confidence values as an additional parameter alongside depth values. This allows the system to encode both the depth information and the reliability of that information in a unified framework, transforming the loss function to incorporate these confidence parameters
2Quantity of substance
If ground truth depth map with filled empty values is used for training, then the completeness of training data is improved, but the reliability of training labels deteriorates
Solution Approach 1:
The patent implements feedback by incorporating confidence information into the loss function during training. The loss function uses confidence values to weight the error terms, providing feedback to the model about which predictions are more critical, thereby guiding the learning process to prioritize accurate regions
Solution Approach 2:
The patent performs preliminary action by pre-processing the ground truth depth map to generate confidence values before training begins. This preliminary confidence map is then used throughout the training process to guide the optimization, allowing the system to prepare reliability information in advance rather than computing it during training
Data Source
AI summary
A method and apparatus for estimating depth information of an image are disclosed. The depth information estimation method of the image includes providing a confidence map for a ground truth depth map; and learning a depth information estimation model for estimating depth information of an image based on the ground truth depth map and the confidence map.