MORDA: A Synthetic Dataset to Facilitate Adaptation of Object Detectors to Unseen Real-target Domain While Preserving Performance on Real-source Domain

Hojun Lim, Heecheol Yoo, Jinwoo Lee, Seungmin Jeon, Hyeongseok Jeon

Keywords: Synthedic-data-assisted domain adaptation, perception, simulation, digital twin

TL;DR: We constructed a novel synthetic dataset, MORDA, that could enable perception models to adapt unforeseen driving scenes without performance degradation on the trained domain.

Abstract: In our accepted work at IEEE ICRA 2025 (https://ieeexplore.ieee.org/document/11127978), we aim to alleviate an indispensable dataset-creation cost in deep learning-based training pipelines for autonomous vehicles (AVs) by leveraging simulation environments. As a concrete demonstration of our method, we create a novel synthetic dataset, MORDA: Mixture Of Real-domain characteristics for synthetic-data-assisted Domain Adaptation. MORDA takes nuScenes and South Korea as source and target domains, respectively. In our extensive experiments, MORDA enables 2D and 3D object detectors to adapt unforeseen driving scenes of South Korea, reporting notable mAP gains, without performance degradation on their source domain, nuScenes.

Serve As Reviewer: ~Hojun_Lim1

Submission Number: 8