Debiased Source-Free Object Detection (DSOD)

• The DSOD framework reduces source bias by employing category-aware adaptive thresholds and low-confidence proposal mining, markedly improving recall for minority classes.
• It integrates semantically rich priors from vision foundation models through dual-source pseudo-label fusion and semantic-aware regularization to enhance cross-domain detection.
• Experimental evaluations demonstrate that DSOD outperforms standard SFOD methods by 4–8 mAP, significantly narrowing the gap to fully supervised performance.

Debiased Source-Free Object Detection (DSOD) is an advanced framework for adapting object detectors from a labeled source domain to an unlabeled target domain, relying solely on the distributional knowledge encoded in the source-pretrained model and without access to any source data during adaptation. The central focus of DSOD approaches is the mitigation of bias—particularly source bias, context/class imbalance, and easy-positive/majority-class skew—that arises in standard Source-Free Object Detection (SFOD) pipelines. Recent advances integrate semantically rich priors from large-scale vision foundation models (VFMs), refined pseudo-label generation protocols, class-relation aware loss shaping, and proposal mining to achieve robust, unbiased cross-domain detection performance.

1. Problem Definition and Motivation

Standard SFOD methods start with a source-pretrained detector $\theta_s$ trained on labeled data $D_s = \{(x^i_s, y^i_s)\}_{i=1}^{N_s}$ and seek to adapt it to an unlabeled target domain $D_t = \{x^i_t\}_{i=1}^{N_t}$ by way of teacher-student or self-training paradigms. However, when adapting exclusively on unlabeled target data, such models suffer from source bias: learned features and decision boundaries are overfit to source-specific distributions or majority classes. Pseudo-labels produced by the teacher on the target domain are systematically biased—overconfident on dominant/majority classes and high-confidence, easily detected instances, while ignoring rare objects, small-scale, or domain-shifted patterns. This results in high false negative rates, error accumulation during self-training, and degraded recall and mAP, especially for minority categories and difficult samples (Li et al., 2020, Ashraf et al., 21 Apr 2025, Cai et al., 19 Jan 2026, Yoon et al., 2024, Zhang et al., 2023).

The objective of DSOD is to construct adaptation and learning mechanisms—across pseudo-labeling, augmentation, feature alignment, and regularization—that explicitly compensate for these biases, thereby improving generalization, transferability, and per-class balance in cross-domain detection.

2. Pseudo-Label Debiasing Protocols

2.1 Confidence Thresholding and Adaptive Thresholds

Early works leveraged a global confidence threshold $h$ for pseudo-label generation. However, setting a fixed threshold yields severe class imbalance: dominant classes (e.g., “car”) are overrepresented while rare or hard classes are filtered out, resulting in poor recall (Li et al., 2020, Zhang et al., 2023). The Refined Pseudo-Labeling (RPL) framework addresses this by employing a Category-Aware Adaptive Threshold Estimation (CATE) procedure: for each class $i$, a distinct threshold $\delta_i$ is dynamically estimated based on the target batch’s confidence distribution, allocating lower $\delta_i$ to under-represented classes and higher $\delta_i$ to dominant ones. This rebalances the number of pseudo-labeled examples per class, directly combating majority bias (Zhang et al., 2023).

2.2 Localization-aware Assignment

Not all high-confidence boxes possess accurate localization. RPL further separates thresholded pseudo-labels into "certain" (well-localized, high mean IoU with proposal cluster) and "uncertain" cases via a localization proxy. Regression losses are only computed for the "certain" set, while the "uncertain" set contributes only classification loss via probability alignment, avoiding regression to noisy coordinates (Zhang et al., 2023).

2.3 Low-Confidence Proposal Mining

Most standard approaches discard low-confidence predictions, thereby ignoring hard positives (small or rare objects with low detection scores). The Low-confidence Pseudo Label Distillation (LPLD) strategy mines proposals from the RPN that fall below the HPL threshold and do not overlap with accepted pseudo-labels. These are further filtered (background score, class-distribution amplification), then supervised via KL-divergence on soft class distributions, but without bounding box regression. Feature similarity between teacher and student proposals is used to weight the loss, focusing on those with strong evidence of foreground semantics (Yoon et al., 2024). This mechanism systematically reduces the false negative rate and improves recall for hard classes.

3. Feature-level Debiasing and Foundation Model Integration

3.1 Vision Foundation Model (VFM)-assisted Alignment

VFMs (e.g., DINOv2, CLIP) serve as external, unbiased priors due to their exposure to massive, heterogeneous, and less source-skewed distributions (Cai et al., 19 Jan 2026, Yao et al., 10 Nov 2025). DSOD frameworks inject VFM features at multiple points:

• Unified Feature Injection (UFI): Frozen VFM features are projected and fused at multiple CNN backbone levels, with fusion strength modulated by a Domain-aware Adaptive Weighting (DAAW) stability criterion. This preserves complementary semantics, improves robustness to shift, and anchors training away from source-specific spurious correlations (Cai et al., 19 Jan 2026).
• Patch-weighted Global Feature Alignment (PGFA): Patchwise student features are aligned to VFM outputs using patch similarity as a weight, focusing adaptation on domain-invariant regions (Yao et al., 10 Nov 2025).
• Prototype-based Instance Feature Alignment (PIFA): Instance-level student RoI features are aligned to momentum-updated class prototypes extracted from the VFM backbone via contrastive (InfoNCE) loss, enhancing instance-level discriminability and reducing source-induced bias (Yao et al., 10 Nov 2025).

3.2 Dual-source Pseudo-label Fusion

Prediction fusion schemes such as Dual-source Enhanced Pseudo-label Fusion (DEPF) merge teacher and VFM detector predictions via an entropy-based weighting within proposal clusters, yielding pseudo-labels that are less biased and less overconfident, especially under domain shift (Yao et al., 10 Nov 2025).

3.3 Semantic-aware Feature Regularization

DSOD leverages Semantic-aware Feature Regularization (SAFR) to enforce consistency between CNN pathway and VFM features, focusing the MSE loss on foreground regions as determined by pseudo-box heatmaps. This constrains the student to learn semantically meaningful features that are less prone to memorizing source-domain spurious patterns (Cai et al., 19 Jan 2026).

3.4 VFM-free Knowledge Distillation

For deployment efficiency, DSOD distills knowledge into a VFM-free student model using a dual-teacher protocol (EMA teacher and frozen VFM-guided teacher), preserving most of the performance benefit at lower inference cost (Cai et al., 19 Jan 2026).

4. Debiasing via Class Context Modeling and Loss Shaping

4.1 Relation Contextual Module (RCM) and Class-relation Augmentation

To tackle context/class imbalance and mode collapse, the Grounded Teacher (GT) framework models inter-class confusion as a global context matrix $\mathbf{R}$, dynamically updated during training. $\mathbf{R}$ is used to identify majority and minority classes and guides MixUp instance-level augmentation: minority classes are specifically amplified by mixing with classes most likely to cause confusion, based on $\mathbf{R}$. Majority class augmentation is biased toward cross-class blends, reducing class domination (Ashraf et al., 21 Apr 2025).

4.2 Semantic-aware Weighted Loss

RCM informs a semantic-aware, reweighted loss. Each student classification error is weighted by its per-class and cross-class confusion probability, discouraging early convergence on majority/easy classes and forcing the model to allocate representational capacity to underrepresented/“hard” categories. The regularization term $\lambda_\ell$ normalizes these weights to prevent instability (Ashraf et al., 21 Apr 2025).

4.3 Expert Foundational Branch

A frozen VFM branch provides additional pseudo-supervision on bounding boxes and class labels. The student's predictions are aligned (via consistency and regression losses) with expert-derived proposals, anchoring adaptation and mitigating teacher drift or mode collapse (Ashraf et al., 21 Apr 2025).

5. Experimental Results and Comparative Performance

DSOD methods are evaluated across a spectrum of challenging benchmarks spanning weather (Normal$\rightarrow$Foggy), synthetic-to-real, cross-scene (Cityscapes$\rightarrow$BDD100K), and artistic style transfer (PascalVOC$\rightarrow$Clipart, Watercolor).

<table> <thead> <tr> <th>Method</th> <th>Cityscapes→Foggy (mAP)</th> <th>Sim10k→Cityscapes (car AP)</th> </tr> </thead> <tbody> <tr> <td>Source only</td> <td\>25.2–29.5</td> <td\>32.0–33.7</td> </tr> <tr> <td>Mean Teacher (SFOD baseline)</td> <td\>34.3–42.3</td> <td\>39.7–42.3</td> </tr> <tr> <td>DSOD (VFM-assisted)</td> <td\>47.1–48.1</td> <td\>61.4–67.4</td> </tr> <tr> <td>Grounded Teacher (class-relation)</td> <td\>50.8</td> <td>—</td> </tr> <tr> <td>Low-conf. Pseudo Label Distillation</td> <td\>40.4</td> <td\>49.4</td> </tr> <tr> <td>RPL (CATE+LPLA)</td> <td\>40.2</td> <td\>50.1</td> </tr> </tbody> </table>

Key observations from the data:

• DSOD with VFM integration consistently outperforms prior SFOD and UDA baselines by +4–8 mAP, approaching fully supervised “oracle” performance in some scenarios (Cai et al., 19 Jan 2026, Yao et al., 10 Nov 2025).
• Grounded Teacher elevates minority-class mAP by 3–5 points and narrows the mAP gap to the oracle, especially on minority and hard classes (Ashraf et al., 21 Apr 2025).
• Low-confidence proposal mining reduces minor-class false negatives and boosts small-object recall by 6–10 mAP (Yoon et al., 2024).
• Adaptive thresholding and localization-aware assignment, as in RPL, provide substantial improvements over fixed-threshold baselines, especially in long-tailed settings (Zhang et al., 2023).

6. Discussions: Limitations, Failure Modes, and Future Directions

Despite significant advances, several limitations remain:

• VFM-based methods incur substantial computational overhead during adaptation. This motivates the use of distillation or compressed backbones for deployment (Cai et al., 19 Jan 2026).
• Failure of both teacher and VFM on rare, severely shifted instances can leave some FNs uncorrected; entropy-aware fusion and further confidence calibration are active research areas (Yao et al., 10 Nov 2025).
• Current methods largely focus on two-stage detectors (Faster R-CNN, Deformable DETR); extension to dense one-stage architectures is ongoing (Yoon et al., 2024).
• The design of augmentation and curriculum strategies that further exploit estimated per-class noise, spatial co-occurrence, or self-supervised structure may further enhance minority-class performance (Ashraf et al., 21 Apr 2025).
• Dynamic scheduling of confidence thresholds, loss coefficients, or curriculum complexity, as well as leveraging self-supervised or multimodal foundation models, represent promising future research directions.

7. Conclusion

Debiased Source-Free Object Detection comprises a family of adaptation algorithms that integrate context-aware pseudo-labeling, proposal mining, class-relational data augmentation, semantic-feature regularization, and external VFM priors to directly address the limitations of standard mean-teacher or pseudo-label self-training under severe class and domain bias. Through principled module design and systematic evaluation—including CATE, LPLA, LPLD, RCM, SAFR, UFI, and PGFA/PIFA/DEPF fusion—these frameworks achieve superior adaptation without access to source data, robustly improving recall on rare classes, minimizing error accumulation, and establishing new state-of-the-art results in cross-domain object detection (Li et al., 2020, Zhang et al., 2023, Yoon et al., 2024, Ashraf et al., 21 Apr 2025, Yao et al., 10 Nov 2025, Cai et al., 19 Jan 2026).