CVPR Findings2026
Once for All: An End-to-End Paradigm for VLM-Based Domain-Generalized Object Detection
Peng Zhang, Xiang Yuan, Cong Li, Junwei Han, Gong Cheng
Abstract
Vision-language models (VLMs), empowered by their crossmodal alignment, have been increasingly exploited for domain generalization. Despite this capability, adaptation for task-specific applications (e.g., object detection) necessitates a typical stage-by-stage training scheme to address architecture mismatch and facilitate domain generalization.
To enable an end-to-end domain-generalized object detection training paradigm, we propose a domain-orthogonal style learning (DoSL) framework that innovatively moves beyond reliance on domain-specific style learning by unifying feature distribution. First, a couple of style biases are learned to normalize the features diversified by perturbing the style statistics, thereby alleviating the style discrepancy between the source domain and the synthesized domains. Furthermore, a domain-orthogonal learning loss is tailored to facilitate style normalization by enforcing orthogonality between the image embedding and all domain-oriented text embeddings, thereby yielding a unified domain-invariant feature distribution. For the first time, DoSL simplifies the optimization pipeline by training once yet consistently generalizes across all target domains without relying on a special prompt. Ultimately, extensive experiments on several domain-generalized object detection benchmarks validate that DoSL achieves SOTA performance and consistently surpasses all competitors by evident margins. The code is available at: https:// github.com/ zhangpeng2001/ dosl.git.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
- Segment Anything cites Learning Transferable Visual Models From Natural Language Supervision
- BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation cites Learning Transferable Visual Models From Natural Language Supervision
- Deformable DETR: Deformable Transformers for End-to-End Object Detection cites FCOS: Fully Convolutional One-Stage Object Detection
Builds on36
- Learning Transferable Visual Models From Natural Language SupervisionAlec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh et al.ICML 2021 · 47,906 citations
- Segment AnythingAlexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao et al.ICCV 2023 · 13,211 citations
- Deformable DETR: Deformable Transformers for End-to-End Object DetectionXizhou Zhu, Weijie Su, Lewei Lu, Bin Li et al.ICLR 2021 · 7,353 citations
- BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and GenerationJunnan Li, Dongxu Li, Caiming Xiong, Steven C. H. HoiICML 2022 · 6,549 citations
- FCOS: Fully Convolutional One-Stage Object DetectionZhi Tian, Chunhua Shen, Hao Chen, Tong HeICCV 2019 · 6,042 citations
Related papers
- Semantic Guided Feature Disentanglement and Reconstruction for Domain Adaptive Object DetectionXiaowei Zhao, Zhide Liu, Yuqing Ma, Xianglong LiuCVPR Findings 2026 · 1 citation
- Learning Domain-Aware Detection Head with Prompt TuningHaochen Li, Rui Zhang, Hantao Yao, Xinkai Song et al.NeurIPS 2023 · 40 citations
- DA-Ada: Learning Domain-Aware Adapter for Domain Adaptive Object DetectionHaochen Li, Rui Zhang, Hantao Yao, Xin Zhang et al.NeurIPS 2024 · 20 citations
- E2E-VLP: End-to-End Vision-Language Pre-training Enhanced by Visual LearningHaiyang Xu, Ming Yan, Chenliang Li, Bin Bi et al.ACL 2021
- Boosting Single-Domain Generalized Object Detection via Vision-Language Knowledge InteractionXiaoran Xu, Jiangang Yang, Wenyue Chong, Wenhui Shi et al.ACM MM 2025 · 2 citations
