Prior-ESM: Enhancing Multicamera 3-D Object Detection With Embedded Semantic Priors and Instance-Level Data Augmentation
IEEE Sensors Journal · 15.04.2026 · 26 (8) · 12806-12817
DOI: 10.1109/jsen.2026.3673602 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Sensors Journal, erschienen am 15.04.2026. Verfasst von Hangwen Zhang und 5 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 52 Literaturangaben. Gefördert von: Outstanding Youth Science Foundation Project (Overseas) of National Natural Science Foundation of China.
Abstract (Originalsprache)
End-to-end 3-D object detection methods based on multicamera setups have recently gained attention for their simple yet effective architectures. However, these approaches suffer from two major limitations. First, they often struggle to capture spatial relationships efficiently. Second, their performance depends heavily on large amounts of well-annotated and distributed data. To address these challenges, we propose Prior-ESM, which integrates embedded semantic maps (ESMs) as priors into the detection queries and uses deformable attention to guide these queries toward features surrounding reference points. By embedding implicit, spatially aware priors into the decoder queries, our method enhances the model’s ability to infer geometric relationships. Moreover, an instance-level data augmentation technique based on the segment anything model2 (SAM2) is introduced. Unlike conventional 2-D image augmentations or multimodal 3-D procedures, our method uses singlemodality images from multiview inputs to enhance datasets within the 3-D spatial domain. By performing data augmentation offline, we mitigate class imbalance without adding computational overhead during training. Prior-ESM demonstrates the state-of-the-art performance on the nuScenes dataset, achieving 48.9% mAP over nontemporal methods and 56.2% mAP over temporal methods. Furthermore, it attains a superior performance of 23.7% mAP on the challenging Argoverse 2 sensor dataset.
Förderung
- Outstanding Youth Science Foundation Project (Overseas) of National Natural Science Foundation of China
Zitieren
APA
Zhang, H., Huang, Y., Liu, Z., Hu, C., Li, G., & Zhang, X. (2026). Prior-ESM: Enhancing Multicamera 3-D Object Detection With Embedded Semantic Priors and Instance-Level Data Augmentation. IEEE Sensors Journal, 26(8), 12806-12817. https://doi.org/10.1109/jsen.2026.3673602
Vancouver
Zhang H, Huang Y, Liu Z, Hu C, Li G, Zhang X. Prior-ESM: Enhancing Multicamera 3-D Object Detection With Embedded Semantic Priors and Instance-Level Data Augmentation. IEEE Sensors Journal. 2026;26(8):12806-12817. doi:10.1109/jsen.2026.3673602
BibTeX
@article{Zhang2026,
author = {Zhang, Hangwen and Huang, Yonghui and Liu, Zhi and Hu, Chuan and Li, Guofa and Zhang, Xi},
title = {Prior-ESM: Enhancing Multicamera 3-D Object Detection With Embedded Semantic Priors and Instance-Level Data Augmentation},
journal = {IEEE Sensors Journal},
year = {2026},
volume = {26},
number = {8},
pages = {12806--12817},
doi = {10.1109/jsen.2026.3673602}
}
EUVIMED