Sensor Frame Annotation by Condition-Based Retraining
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for annotating sensor data, particularly image acquisition data for autonomous driving, require significant human effort and time-consuming quality checks, making large-scale annotation projects infeasible with human labor alone, and existing automation approaches are inefficient in maintaining high annotation quality.
Innovation Solution
A method involving grouping sensor data frames by environmental conditions, using neural networks for initial annotation, selecting samples for quality measurement, retraining when necessary, and focusing retraining on frames with sub-threshold quality to improve annotation accuracy, thereby reducing manual effort and computing power.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If neural networks are used to automate annotation of sensor data, then productivity increases and manual effort decreases, but maintaining high annotation quality requires time-consuming quality checks that increase the linear relationship between project volume and workload
Solution Approach 1:
The patent segments the annotation quality assurance process by dividing sensor data frames into packets based on environmental conditions, and further segments quality checks by applying them selectively only to packets with sub-threshold quality measures rather than uniformly to all data. This segmentation reduces overall quality check time while maintaining annotation productivity.
Solution Approach 2:
The patent applies different quality assurance levels to different segments of data based on their specific characteristics. Packets with above-threshold quality measures skip detailed quality checks, while only packets with sub-threshold measures undergo retraining and re-evaluation. This local quality approach optimizes the balance between productivity and quality maintenance.
2Reliability
If comprehensive quality checks are applied to all annotations to ensure high quality, then annotation quality is maintained, but the workload increases linearly with project volume requiring more human resources
Solution Approach 1:
The patent segments the quality assurance workload by dividing all annotations into multiple packets based on environmental conditions, then applies quality checks only to specific segments (packets with sub-threshold measures) rather than uniformly to all segments. This reduces the complexity and human resources required while maintaining overall annotation quality.
Solution Approach 2:
Different quality assurance measures are applied locally to different packets based on their quality measures. High-quality packets receive minimal or no additional quality checks, while low-quality packets receive focused retraining and re-evaluation. This local approach maintains reliability without proportionally increasing workload complexity.
3Measurement precision
If retraining is performed on all training data to improve annotation quality, then annotation accuracy improves, but computing power and time requirements increase significantly
Solution Approach 1:
The patent extracts and identifies specific packets with sub-threshold quality measures from the overall training data, then applies retraining only to these problematic segments rather than retraining the entire dataset. This extraction approach improves annotation accuracy for critical cases while significantly reducing computing power consumption.
Solution Approach 2:
Retraining is applied locally only to packets demonstrating sub-threshold quality measures rather than uniformly to all training data. This targeted local retraining improves annotation accuracy where needed while minimizing the energy and computing power requirements compared to comprehensive retraining.
4Productivity
If the number of labelers is increased to complete annotation projects within shorter timeframes, then productivity increases, but the cost and complexity of managing human resources increases
Solution Approach 1:
The patent implements a self-service quality assurance mechanism where the system automatically identifies packets with sub-threshold quality measures and triggers targeted retraining without requiring human intervention for quality checks. This automation reduces productivity dependency on human labelers while simplifying project management complexity.
Data Source
Figure 1~2
Figure 3~5
Figure 6
AI summary
The invention relates to a computer-implemented method for automatically annotating sensor data frames such as video frames or audio frames. Received frames are grouped into a multiplicity of packets on the basis of at least one conditional attribute which relates to a surroundings condition that existed when the frame was recorded. A first packet corresponding to a specific value range of the conditional attribute is annotated using a neural network. The computer determines a quality level for the annotations on the basis of a first sample frame. If the quality level for at least one frame is below a predefined threshold value, the neural network is retrained on the basis of corrected annotations for the first sample. If the quality level is above the predefined threshold value, the annotated frames are exported. The invention also relates to a nonvolatile computer-readable medium and to a computer system.