Learning Device Depth Map Variance Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The consistency between depth maps generated by LiDAR sensors and images from cameras is often low due to calibration errors and relative motion, leading to inaccurate training of depth map estimation networks and requiring extensive time and cost to select good-quality depth maps for training.
Innovation Solution
A learning device that obtains depth maps and images, calculates variance estimation information indicating the variance between them, back-propagates variance loss to update parameters, and determines suitability for training, thereby improving the accuracy of depth maps and reducing the time and cost of constructing training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If depth maps are generated using LiDAR sensors and cameras, then depth information can be obtained for training, but consistency between the depth map and image is low due to calibration errors and relative motion
Solution Approach 1:
The patent implements a feedback mechanism by calculating variance between depth maps and images, then back-propagating variance loss to update parameters of the depth map generation model. This closed-loop feedback system continuously improves consistency by using the calculated variance as a corrective signal, transforming the inconsistency problem into an optimizable loss function that guides parameter adjustments.
Solution Approach 2:
The patent applies parameter changes by updating the parameters of the depth map generation model based on back-propagated variance loss. This involves adjusting calibration parameters, transformation matrices, or model weights to minimize the variance between predicted depth maps and ground truth images, thereby improving consistency through systematic parameter optimization.
2Reliability
If inspectors manually select good-quality depth maps, then training data quality can be maintained, but huge amounts of time and cost are consumed
Solution Approach 1:
The patent enables self-service by implementing an automated quality assessment system that calculates variance between depth maps and images to determine training data suitability. This eliminates the need for manual inspector selection, as the system autonomously evaluates and selects high-quality training data based on quantitative variance metrics, significantly reducing time and cost while maintaining data quality.
Solution Approach 2:
The patent replaces the mechanical manual inspection process with an automated computational system. Instead of human inspectors visually evaluating depth map quality, the system uses algorithmic variance calculation and back-propagation to objectively assess and select training data, substituting manual labor with automated image processing and machine learning techniques.
3Productivity
If depth maps with low consistency are used for training, then training data can be constructed quickly, but accuracy of the depth map estimation network is low
Solution Approach 1:
The patent applies preliminary action by pre-calculating variance between depth maps and images before training to identify and select high-quality training data. This preliminary quality assessment ensures that only depth maps with acceptable consistency are used for training, preventing the network from learning from poor-quality data while maintaining efficient training data construction through automated filtering.
Data Source
AI summary
A learning device is introduced. The device may comprise a processor, and memory storing instructions that, when executed by the processor, may cause the device to obtain at least one first depth map based on at least one piece of cloud data associated with surrounding environment information, and at least one first image associated with the at least one first depth map, determine, based on the at least one first depth map and the at least one first image, variance estimation information indicating a variance between the at least one first depth map and the at least one first image, back-propagate a variance loss based on the first variance estimation information, and variance ground truth (GT) information associated with the first variance estimation information, and update, based on the back-propagated variance loss, a parameter associated with determining the first variance estimation information.