Pseudo-ground-truth Generation from Timestamp Supervision

Simple SummaryContent extracted from patent full text and abstract with AI.

This invention provides a computer-implemented method for generating 'pseudo-ground-truth' data to train machine learning models for segmenting or localizing patterns within sequences of data, such as images, video, audio, or other signals. By using sparse 'timestamp' annotations (where a frame and its pattern are labeled) and dividing the data between timestamps into labeled and neutral regions, the method generates optimized training labels that are robust even when not all patterns are annotated. This process helps produce high-quality training data without needing full, labor-intensive manual annotation of every frame.

Use CasesContent extracted from patent full text and abstract with AI.

• Training video action segmentation models with only sparse timestamp annotations
• Developing audio segmentation systems with fewer labeled examples
• Improving machine learning for medical signal processing (e.g., ECG/EEG signal segmentation)
• Automated segmentation and localization in industrial monitoring (e.g., manufacturing processes)
• Efficient generation of training data for autonomous driving perception systems
• Enhancing annotation efficiency in large-scale image or video datasets for research or commercial use

BenefitsContent extracted from patent full text and abstract with AI.

• Greatly reduces the need for dense manual annotation, saving time and cost
• Provides robustness to missing or incomplete annotation segments
• Improves the quality and reliability of machine learning model training with weak supervision
• Can be applied to various types of data (video, audio, image sequences, sensor signals)
• Offers flexibility by allowing ignored regions where annotation certainty is low or missing
• Enables better performance in tasks like action segmentation and localization, even with sparse data

Technical Classifications (CPCs)

Inventors & Applicants

Patent Abstract

A method for generating pseudo-ground-truth data for training a machine learning model to segment or localize patterns in a sequence of frames, comprising:- providing a sequence of frames, and timestamps each including identification of a frame and a pattern annotation thereof;- defining at least one subset comprising the frames located between a first timestamp (pi) and a consecutive second timestamp (pi+1);- initially assigning the frames of the subset to a first region (55) starting from the first timestamp (pi), a second region (59) starting from the second timestamp (pi+1) or a neutral region (57) joining the first region (55) to the second region (59);- obtaining pattern estimations for each frame of the subset;- optimizing boundaries between the regions (55, 57, 59) based on a compromise between the size of the neutral region (57) and a match between the pattern estimations of the frames and the regions to which the frames are assigned;- outputting pseudo-ground-truth data comprising labels for the first and second regions.