CD-FKD: Cross-Domain Feature Knowledge Distillation for Robust Single-Domain Generalization in Object Detection

Junseok Lee1,2, Sungho Shin3, Seongju Lee1, Kyoobin Lee1
1Gwangju Institute of Science and Technology (GIST), 2LG Electronics, 3NAVER Corp.
Paper arXiv Video Code

Teaser Image / Video Placeholder

CD-FKD improves robust object detection in unseen domains by distilling feature knowledge.

Abstract

Single-Domain Generalization (SDG) in Object Detection aims to train a detector on a single source domain that generalizes well to unseen target domains. Existing methods often rely on image-level augmentations or feature alignment, which may not fully capture the underlying invariant features required for robust detection. In this paper, we propose Cross-Domain Feature Knowledge Distillation (CD-FKD), a novel framework that improves generalization by distilling feature-level knowledge. We introduce a teacher-student architecture where the teacher learns from augmented views of the source domain, and the student learns to reconstruct these features from the original view, effectively learning domain-invariant representations. Extensive experiments on multiple benchmark datasets demonstrate that CD-FKD significantly outperforms state-of-the-art SDG methods in various domain shift scenarios.

Method

Detailed explanation of the proposed method goes here. You can add diagrams and text describing the architecture, loss functions, and training procedure.

Method Diagram Placeholder

Results

Quantitative and qualitative results demonstrating the effectiveness of CD-FKD.

Results Table / Figure Placeholder

BibTeX

@inproceedings{lee2026cdfkd,
  author    = {Lee, Junseok and Shin, Sungho and Lee, Seongju and Lee, Kyoobin},
  title     = {CD-FKD: Cross-Domain Feature Knowledge Distillation for Robust Single-Domain Generalization in Object Detection},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  year      = {2026},
}