Training device, training method, object recognition device, object recognition method, program, and machine learning model

WO2025254156A1PCT designated stage Publication Date: 2025-12-11PANASONIC INTELLECTUAL PROPERTY MANAGEMENT CO LTD
0 Cites 0 Cited by

Patent Information

Application Number
PCT/JP2025/020258
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2024-06-05
Filing Date
2025-06-04
Publication Date
2025-12-11

AI Technical Summary

Technical Problem

Existing machine learning models, particularly Vision Transformers, require large amounts of real image datasets for pre-training, which pose privacy issues and label biases, and synthetic datasets like VisualAtom and FDSL may not achieve the same recognition accuracy as real images.

Method used

A training technique using Formula-driven Supervised Learning (FDSL) data and Computer Graphics (CG) data to generate a pre-trained model that learns contour information and object-likeness, allowing for accurate pre-training without real images.

Benefits of technology

The combined use of FDSL and CG data enables the generation of a pre-trained model that captures the entire object's features accurately, extending the region of interest beyond what FDSL alone can achieve, and fine-tuning this model with actual images results in a highly accurate machine learning model for specific tasks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure JP2025020258_11122025_PF_FP_ABST
    Figure JP2025020258_11122025_PF_FP_ABST
Patent Text Reader

Abstract

One aspect of the present disclosure relates to a training device comprising: an acquisition unit that acquires formula-driven supervised learning (FDSL) data representing a target object and computer graphics (CG) data; and a training unit that trains a machine learning model to be trained using the FDSL data and the CG data.
Need to check novelty before this filing date? Find Prior Art