Verfahren zur Bestimmung von 3D-Belegungsvorhersagen einer Szene unter Verwendung eines 3D-Belegungsvorhersagenetzwerks

Details

Veröffentlichungs-Nr.

EP4800657

Anmeldetag

26. Februar 2025

Veröffentlichung

2. September 2026

Rechtsraum

EP

Abstract

A method for training a 3D occupancy prediction network , said method comprising a. Obtaining (S1) a plurality of input images (Ik) of a scene, b. Extracting (S3), a relative depth information for 3D points of a target image (Itarget), c. Obtaining (S4), an absolute depth information (Dtarget) of said target image (Itarget), d. Obtaining (S5) a reconstructed target image using said absolute depth information (Dtarget), e. Obtaining (S6), semantic information of pixels of said target image (Itarget) and semantic information of pixels of said reconstructed target image (Îtarget), f. Selecting (S7) pixels (u') of said target image whose semantic information is the same in said target image (Itarget) and in said reconstructed target image (Îtarget), g. optimizing (S8) a semantic loss ( L sem − refine <img class="EMIRef" id="6f4b85ad-6654-4121-9a00-e379a7f9b3e4-ia01" /> ) function as a binary cross entropy applied on patches ( S i ′ u i ′ <img class="EMIRef" id="6f4b85ad-6654-4121-9a00-e379a7f9b3e4-ia02" /> ) and reconstructed patches ( S ^ i ′ u i ′ <img class="EMIRef" id="6f4b85ad-6654-4121-9a00-e379a7f9b3e4-ia03" /> ) of said selected pixels, h. optimizing (S9) a depth supervision loss ( L depth <img class="EMIRef" id="6f4b85ad-6654-4121-9a00-e379a7f9b3e4-ia04" /> ) function

Anmelder

Firmen

TOYOTA JIDOSHA KABUSHIKI KAISHA
Max-Planck-Gesellschaft zur Förderung der Wissenschaften e.V.

Land

🇯🇵 Japan