| Title |
MLDA-DETR: Multi-level Domain Alignment for Bridging the Synthetic-to-real Gap in DETR Object Detection |
| Authors |
용윤정(Yunjeong Yong) ; 박재우(Jaewoo Park) |
| DOI |
https://doi.org/10.5573/ieie.2026.63.3.108 |
| Keywords |
Synthetic-to-real domain adaptation; Multi-level domain alignment; DETR; Prototype-based query alignment; Object detection |
| Abstract |
Synthetic data offers a viable alternative for training object detectors where real-world data is scarce. However, stylistic discrepancies impair both backbone features and decoder queries in DETR models. To address this, we propose MLDA-DETR, a multi-level domain alignment framework. Our approach decomposes domain shift into feature and query levels, aligning them via end-to-end optimization using statistical/textural alignment and Class-aware Prototype Alignment (CAPA). Experiments on the Roboflow Anime-to-Real dataset yielded an mAP of 58.2, approaching real-only performance. Furthermore, cross-domain evaluations on Foggy Cityscapes and BDD100k demonstrate competitive generalization without target domain data. |