Fusing depth embeddings for monocular 3D object detection

doi:10.1016/j.jvcir.2025.104555

Fusing depth embeddings for monocular 3D object detection

2025-08-07

0

PRE

AI

DOI：10.1016/j.jvcir.2025.104555

Original

Original request for help

Save

Abstract

En 中文

• A novel depth-assisted monocular 3D object detection method is proposed. The depth map, generated by a pre-trained monocular depth estimator, is embedded into a transformer-based detection framework as learnable depth embeddings, providing effective guidance for 3D object detection. • A lightweight keypoint detection module is designed to predict the 3D projective centers of objects. The features at these locations are then sampled to initialize object queries, effectively reducing the learning complexity of the transformer decoder.

Keywords:

monocular 3D object detection

depth estimation

transformer-based framework

keypoint detection

object query initialization

Fusing depth embeddings for monocular 3D object detection

Abstract

Journal

Organization