Training system, training method, and imaging device

The training system enhances the accuracy and efficiency of three-dimensional reconstruction models by selecting optimal image types and models based on scene information, addressing issues of image quality and processing inefficiencies in existing AI models like NeRF and 3D-GS.

WO2026133675A1PCT designated stage Publication Date: 2026-06-25SONY SEMICON SOLUTIONS CORP
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
SONY SEMICON SOLUTIONS CORP
Filing Date
2025-09-29
Publication Date
2026-06-25

AI Technical Summary

Technical Problem

Existing three-dimensional reconstruction models using AI, such as Neural Radiance Fields (NeRF) and 3D-Gaussian Splatting (3D-GS), face issues with deteriorated image color quality when event information is used for error calculation in non-blurred input images, and inefficient training processing.

Method used

A training system that includes an imaging device with multiple sensors (gradation, event, and depth) and a server device for training, which acquires scene information to select appropriate image types and models based on the scene, optimizing training efficiency and accuracy by selecting the right type of captured images and models for training.

Benefits of technology

Improves the generation accuracy and efficiency of free viewpoint images by appropriately training the three-dimensional reconstruction model according to the imaging scene, preventing unnecessary image capture and optimizing training processing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure JP2025034354_25062026_PF_FP_ABST
    Figure JP2025034354_25062026_PF_FP_ABST
Patent Text Reader

Abstract

A training system includes: an imaging unit having an image sensor configured to capture an image for training a three-dimensional reconstruction model and is configured to be able to acquire a plurality of types of captured images as the image; and circuitry configured to acquire scene information indicating a type of a scene to be imaged by the imaging unit select a type of the three-dimensional reconstruction model and a type of the captured image to be used for the training, on the basis of the scene information previously acquired; and train a three-dimensional reconstruction model of the model type previously selected using a captured image of the type selected.
Need to check novelty before this filing date? Find Prior Art