Vorrichtung und Verfahren zum Aufwärtsskalieren einer Merkmalskarte eines Bildes

Details

Veröffentlichungs-Nr.

EP4826643

Anmeldetag

16. April 2026

Veröffentlichung

21. Oktober 2026

Rechtsraum

EP

IPC

G06T5/60

Abstract

Computer-implemented method for training a first machine learning system, wherein the first machine learning system (60) is configured for determining an output (o) characterizing an upsampling a feature map of an image (x) using the image (x) and a feature map (F) of the image (x) as input, the method comprising the steps of:• Training the first machine learning system (60) in a first stage, wherein the first stage comprises a pretraining of the first machine learning system (60);• Training the first machine learning system (60) in a second stage, wherein the second stage comprises:• Training a second machine learning system (60b) according to the first stage;• Obtaining a training image (x_i) and a feature map (F_i) of the training image (x_i);• Determining a first upsampled feature map (o_a) by providing the training image (x_i) and the feature map (F_i) as input to the first machine learning system (60);• Determining a second upsampled feature map by providing a crop (s_i) of the training image (x_i) and a corresponding crop of the feature map (F_i) as input to the second machine learning system (60b);• Training the first machine learning system (60) based on a second loss function ( ), wherein the second loss function ( ) characterizes a difference between a crop (g_a) of the first upsampled feature map (o_a) and the second upsampled feature map (o_b), wherein the crop (g_a) of the first upsampled feature map (o_a) corresponds to the crop (s_i) of the training image (x_i).

Anmelder

Firma

Robert Bosch GmbH

Land

🇩🇪 Deutschland

Fachgebiete