Detail View

Title
MonoGDA: Monocular 3D Object Detection via Geometry-Aware Depth Alignment
Issued Date
2026-12
Citation
IEEE ROBOTICS AND AUTOMATION LETTERS, v.11, no.12, pp.14372 - 14379
Type
Article
Author Keywords
Object detection ; Modeling ; Geometry ; Measurement ; Distance measurement ; Laser radar ; Learning (artificial intelligence) ; Visualization ; Signal detection ; Depth measurement ; Deep learning for visual perception ; computer vision for transportation ; monocular 3D object detection ; depth estimation ; depth-aware attention
ISSN
2377-3766
Abstract

Inferring 3D information from a single image inherently involves depth ambiguity, as geometric information is compressed during the projection of a 3D scene onto the 2D image plane. To address this challenge, existing monocular 3D detectors leverage object-wise depth to acquire 3D geometric information. This strategy directly provides object-centric metric depth cues derived from 3D bounding boxes to the decoder, thereby reducing depth uncertainty and improving 3D localization performance. However, due to the inherent property of assigning a single depth value to each object region, such approaches cannot sufficiently capture high-level structural cues, which consequently leads to the accumulation of depth bias. In this work, we propose MonoGDA to overcome these limitations. MonoGDA exploits scene-level depth, which captures the global geometric structure of the image, together with object-wise depth, which provides object-level metric constraints, and introduces a unified depth representation that integrates the two depth cues. Furthermore, to effectively model relative depth relationships among objects, we propose a Depth-Aware Attention module that incorporates geometric priors into self-attention. Experimental results show that by leveraging external depth priors, the proposed method achieves state-of-the-art performance among monocular 3D detectors on the KITTI 3D object detection benchmark.

더보기
URI
https://scholar.dgist.ac.kr/handle/20.500.11750/61178
DOI
10.1109/LRA.2026.3746016
Publisher
IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
Show Full Item Record

File Downloads

  • There are no files associated with this item.

공유

qrcode
공유하기

Related Researcher

진용식
Jin, Yongsik진용식

AX Research Group for Robotics

read more

Total Views & Downloads

???jsp.display-item.statistics.view???: , ???jsp.display-item.statistics.download???: