Verfahren und System zur Oberflächenerkennung, Trainingsverfahren und System für das Oberflächenerkennungssystem, Computerprogramm, Computerlesbares Medium

Details

Veröffentlichungs-Nr.

EP4145406

Aktenzeichen

EP21194430

Anmeldetag

1. September 2021

Veröffentlichung

8. März 2023

Rechtsraum

EP

IPC

G06V20/56G06V10/82G06V10/44

Abstract

The surface identification method for identifying pixels representing a certain surface in a scene image (si) comprises:S20) encoding the scene image (si) into a segmentation latent feature (slf);S30) based on the segmentation latent feature (slf), predicting a segmentation (sl,sm) classifying pixels of the scene image into either surface categories (rd) or non-surface categories;S40) based on the segmentation latent feature (slf), calculating a reconstructed scene image (rsi) using a neural network-based reconstruction module (RM), trained to minimize the reconstruction errors (Err) for the surface pixels and to maximize the reconstruction errors (Err) for the non-surface pixels;S50) based on the scene image (si) and the reconstructed scene image (rsi), calculating a reconstruction error map (rem);S60) based on the segmentation (sl,sm) and the reconstruction error map (rem), calculating a surface pixels datum (spm) in which the surface pixels of the scene image (ri) are identified.

Anmelder

Firmen

TOYOTA JIDOSHA KABUSHIKI KAISHA
Czech Technical University in Prague

Land

🇯🇵 Japan

Fachgebiete