acceptodds
Withdrawn from ICLR 2027

Source-Free Object Detection via Multi-Source Knowledge Fusion

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Source-free domain adaptation (SFDA) adapts a pretrained source model to an unlabeled target domain without accessing source data, which is useful when source data cannot be retained or transferred. Multi-source source-free domain adaptation (MSFDA) further exploits complementary knowledge from multiple pretrained source models, but heterogeneous source-target shifts and noisy pseudo labels make multi-source object detection particularly challenging. We propose a multi-source knowledge-fusion framework that addresses these issues through three complementary mechanisms. First, Text-Driven Feature Augmentation (TFA) transforms unlabeled target images toward each source-domain appearance using only coarse semantic descriptions of the source domains, reducing the mismatch seen by the corresponding source experts. Second, the source experts are locally adapted on their source-stylized target views, while an aggregator integrates their knowledge using adaptively learned contribution weights. Third, mutual confidence selection between the aggregator and domain experts reduces the influence of noisy pseudo labels during self-training. Extensive experiments across multiple real-world domain shifts show consistent improvements over existing SFDA and MSFDA baselines under the standard MSFDA protocol.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Source-free domain adaptation, multi-source knowledge-fusion

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ooxNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a multi-source source-free domain adaptation framework for object detection. The method introduces Text-Driven Feature Augmentation (TFA) to reduce source-target discrepancies, adaptive knowledge fusion to integrate multiple source models, and mutual confidence selection to mitigate noisy pseudo labels. Experiments on weather and artistic domain shifts demonstrate improvements over existing SFDA and MSFDA methods.

Critical strengths
  1. Clear and well-organized writing. The manuscript is generally well written, with a clear introduction and a logical presentation of the proposed framework. The motivation and technical components are explained coherently, making the overall methodology relatively easy to follow.

  2. Comprehensive supplementary material. The appendix is relatively rich and provides additional experimental results, implementation details, architectural comparisons, hyperparameter analyses, and qualitative visualizations. These supplementary materials help readers better understand the experimental settings and the behavior of the proposed framework.

Critical weaknesses
  1. The framework illustrations lack clarity and visual quality. The current figures, particularly Figure 1, are not sufficiently clear or visually appealing. The overall architecture is difficult to understand from the illustration alone, and the interactions among TFA, local expert adaptation, adaptive knowledge fusion, and mutual confidence selection are not clearly presented. I recommend substantially redesigning the framework figures with clearer data flows, more readable annotations, and better visual organization so that readers can understand the main technical procedures without relying heavily on the text.

  2. Limited methodological novelty and insufficient justification for TFA. The novelty of the proposed framework appears somewhat limited, as its main components are closely related to existing techniques in text-guided style transfer, model aggregation, and pseudo-label-based self-training. In particular, while recent domain-diversification approaches increasingly explore diffusion-based generation to model complex visual variations, the proposed TFA relies on CLIP-guided feature transformations and subsequent image reconstruction. The advantages of this design over more expressive generative approaches are not sufficiently demonstrated. More importantly, Table 1 shows that introducing TFA improves mAP from only 33.5% to 34.2%, corresponding to a marginal gain of 0.7 percentage points. This limited improvement raises concerns about whether TFA provides a sufficiently meaningful contribution to justify its additional complexity.

Questions for authors
  1. The visualizations do not convincingly demonstrate realistic appearance transformations, and it remains unclear whether alignment in CLIP's semantic embedding space necessarily reduces the visual distribution gap between source and target domains. Can the authors provide quantitative evidence of this alignment and compare TFA with diffusion-based alternatives?

  2. Unlike conventional teacher-student EMA updates, which maintain a temporal relationship between two models, the proposed framework combines parameters from independently trained source-domain models. How can the authors ensure that parameter averaging preserves domain-specific strengths rather than weakening them? What evidence supports the compatibility of these models and the effectiveness of the learned aggregation weights in preventing knowledge conflicts?

Reviewer_xks4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a multi-source knowledge-fusion framework that addresses these issues through three complementary mechanisms. First, Text-Driven Feature Augmentation (TFA) transforms unlabeled target images toward each source-domain appearance using only coarse semantic descriptions of the source domains, thereby reducing the mismatch seen by the corresponding source experts. Extensive experiments show that the proposed method outperforms existing methods.

Critical strengths
  1. The paper targets multi-source-free domain adaptation for object detection (MSFDAOD), where multiple pre-trained source models are available but source data is inaccessible. This setting is well motivated by data-privacy and model-reuse constraints.
  2. The experimental results look good.
Critical weaknesses
  1. The introduction of source-domain text descriptions may violate the source-free assumption. The task of MSFDAOD is defined under the constraint that only pre-trained source models—not source data or source-side information—are accessible. While a pre-trained model is an endogenous, task-given carrier of source knowledge, a hand-written textual description of the source domain is an external, additional source-side prior that is not part of the original task setting. Allowing such descriptions opens the question of where the boundary of "source-free" lies: if a textual style description is permitted, why not source class names, source statistics, or a few source samples? The paper does not discuss this boundary, and the claim of being "source-free" is therefore weakened.

  2. The use of source-domain text descriptions introduces extra human annotation, making comparisons with prior MSFDAOD methods potentially unfair. Baseline methods operate without any source-side textual input; they rely solely on pre-trained models. This paper, in contrast, requires an additional manually written description per source domain to drive TFA. The performance gains may therefore stem from this extra, human-provided source prior rather than from the proposed fusion framework itself.

  3. Some statements in the paper are unnecessarily obscure. For example, the introduction states that "TFA creates source-aligned views of the unlabeled target data from coarse source-domain semantics; local expert adaptation uses these aligned views to retain domain-specific expertise," but it does not explain how this is achieved.

Questions for authors

My major concern is that using text descriptions violates the source-free setting. To make the comparison fair, the authors should either (i) report a variant that does not use text descriptions, or (ii) clarify that the setting is not strictly comparable to prior source-free methods and adjust the claims accordingly.

Reviewer_ryKqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses multi-source-free domain adaptation for object detection. It proposes a framework with two main components. First, a text-driven feature augmentation module uses CLIP to transfer target images toward source-domain styles based on short text descriptions, reducing the style gap between target images and source detectors. Second, a multi-source knowledge fusion framework designates one source model as an aggregator and the rest as domain experts. Experts are updated on style-augmented target data via self-training, the aggregator is updated through EMA from experts, and contribution weights are learned by a small network. Pseudo-label quality is improved through mutual confidence selection between the aggregator and domain experts. Experiments are conducted on DWD, Art, and Foggy-Cityscapes. The method outperforms compared SFDA and MSFDA baselines under [email protected].

Critical strengths
  • The method requires only a short semantic description of each source domain, not source images or style reference images. Table 4 shows that different relevant phrasings yield comparable performance while unrelated text degrades sharply. This is consequential because it makes multi-source-free adaptation more deployable when source data cannot be shared.
  • Evaluation spans multiple datasets, architectures, and tasks. The paper reports results on DWD, Art, and Foggy-Cityscapes, extends experiments to ATSS, YOLOv7, and DETR with ViT, and includes classification and segmentation evaluations in the appendix. The adaptation and inference cost comparison in Table 18 is also useful. This breadth strengthens confidence that the method is not tied to a single detector or dataset.
Critical weaknesses
  • Sec. 4.1 states that all reported results are based on a single run. On DWD, the improvements over the strongest baseline Bi-ATEN are 0.6, 0.2, 0.8, and 3.0 mAP for NC, DR, NR, and DF respectively; on Art, the gains are 1.62, 1.57, and 1.15. Without variance estimates, it is difficult to judge whether these margins are reliable, especially since some baselines are re-implemented. This is consequential because the claim of consistent improvement rests on differences that may be within run-to-run noise.

  • Table 1 shows that from single-source to MS, mAP improves from 31.4 to 33.5; adding TFA improves it to 34.2; adding confidence selection further improves it to 38.4. TFA is a core contribution highlighted in the title and method, yet its standalone gain is much smaller than a relatively standard co-teaching technique. The paper does not provide a factorial decomposition of MS+TFA without CS versus MS+CS without TFA. This directly affects the assessment of the method's core novelty.

  • Table 16 shows that aggregator choice on Night Rainy varies from 17.5 to 20.3 mAP. The paper describes this as choosing a source model whose coarse domain semantics are close to the target, such as weather and time attributes. This requires knowing target-domain attributes in advance. If this is a design choice rather than an oracle, the selection rule should be stated explicitly; if it requires target metadata, the method is less source-free than claimed.

  • In Table 4, Description 1 yields 38.4 mAP and Description 2 yields 36.5 mAP, a gap of 1.9 points. The paper calls them comparable, but 1.9 mAP is not negligible in detection. The paper does not report variance across description samples or state whether 38.4 is the best result after description selection.

Questions for authors
  • Please clarify whether the main results in Tables 2 and 3 are from a single run and whether variance across runs was measured. If variance is small, reporting it would materially strengthen the empirical claims.

  • Does the aggregator selection in Table 16 use target-domain metadata such as weather or time? If so, how is this obtained in real deployment? If not, please explain how the semantically closest source model is identified without target labels.

  • Table 1 reports only cumulative ablations. Please provide complete results for MS+TFA without CS and MS+CS without TFA. In particular, TFA alone contributes only 0.7 mAP while CS contributes 4.2 mAP. Please explain whether this attribution affects the paper's positioning of its core contribution.

  • In Table 4, Description 1 and Description 2 differ by 1.9 mAP, and the paper calls them comparable. Please clarify whether 38.4 is the best result after sampling multiple descriptions, whether more description variants were tested, and whether this gap is acceptable.

  • The paper states that some SFDA and MSFDA baselines are re-implemented. Please list which baselines are re-implemented, whether their hyperparameters were tuned with the same effort as the proposed method, and whether they used the same backbone and training budget.

  • Table 2 uses Day Clear plus the other three domains as sources for each target. Please clarify whether this source combination uses target-domain identity information and how the source model set would be selected in deployment without knowing the target domain.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.