Accepted at ICCV 2025, TITAN introduces a novel framework for Source-Free Domain Adaptive Object Detection (SF-DAOD). In computer vision, adapting a model to a new environment—such as transitioning from sunny streets to foggy conditions—typically requires access to the original training data. TITAN breaks this requirement, achieving state-of-the-art results using only a pre-trained model and unlabeled target data.
TL;DR
TITAN addresses the "model collapse" problem in source-free domain adaptation by partitioning target data into "easy" and "hard" subsets based on prediction variance. By combining global query-token adversarial learning with local token-wise alignment, it achieves significant performance gains—up to +22.7 mAP—across natural and medical imaging benchmarks.
The Challenge of Source-Free Adaptation
Deep learning models are notoriously brittle when faced with domain shift—changes in lighting, weather, or hardware. While Unsupervised Domain Adaptation (UDA) helps, it usually requires the original "source" dataset to be present. In fields like medical imaging or proprietary industrial applications, source data is often unavailable due to privacy or legal constraints. This is the Source-Free (SF) challenge.
Existing SF-DAOD methods often utilize a "student-teacher" framework where a teacher model generates "pseudo-labels" for the student. However, if the domain shift is large, the teacher produces noisy, incorrect labels. The student then learns these errors, leading to a feedback loop of failure known as model collapse.
The TITAN Framework
The authors propose the Target-based Iterative Query-Token Adversarial Network (TITAN) to stabilize this process through two main innovations: target division and query-token alignment.
1. Target Domain Division (Easy vs. Hard)
Instead of treating all target images equally, TITAN uses Monte-Carlo (MC) Dropout to estimate detection variance. Counter-intuitively, the authors discovered that higher detection variance often signals higher recall and greater similarity to the source domain.
As shown in the figure above, this metric allows TITAN to split target data into "easy" (source-similar) and "hard" (source-dissimilar) subsets, enabling a more stable, staged curriculum for adaptation.
2. Query-Token Adversarial Learning
Built on the FocalNet-DINO architecture, TITAN injects a special "query token" into the transformer sequence to handle alignment at two levels:
- Global Alignment: The query token aggregates global scene context and inter-object relationships. A discriminator attempts to distinguish if this token originates from a source-like or target-like image, forcing the model to learn domain-invariant global features.
- Local Alignment: Simultaneously, individual tokens are aligned to handle local textures and styles.
- Cascaded Refinement: This process is applied across multiple layers of the transformer (from shallow to deep) to ensure thorough feature refinement.
Key Results and Insights
TITAN significantly outperformed existing state-of-the-art (SOTA) methods across a comprehensive suite of benchmarks, including Cityscapes, Foggy Cityscapes, and BDD100k.
- Massive mAP Boosts: On natural imaging benchmarks, TITAN improved mAP by +22.7 (C2F), +22.2 (C2B), and +21.1 (S2C) compared to previous SOTA methods.
- Medical Breakthrough: In cross-domain breast cancer detection (RSNA-BSD1K to DDSM), TITAN achieved a recall of 0.78, compared to just 0.25 for the previous best method. This highlights the framework's utility in medical AI, where data silos make source-free adaptation essential.
- Variance as a Compass: The discovery that prediction variance can reliably partition a target domain is a major contribution, providing a principled way to guide the adaptation process.
- The Power of the Query: By condensing domain information into a single Query Token, TITAN makes adversarial alignment more efficient and focused on global scene layout rather than just pixel-level noise.
Limitations and Future Work
While powerful, TITAN's current implementation is tightly integrated with the FocalNet-DINO transformer architecture. Its portability to simpler architectures like Faster R-CNN remains to be explored. Additionally, estimating variance via Monte-Carlo Dropout requires multiple forward passes, increasing the initial computational cost of adaptation. The authors suggest that this query-token approach could eventually be extended to other source-free tasks, such as semantic or instance segmentation.
Summary updated