acceptodds
Withdrawn from ICLR 2027

Single-Domain Generalizable Open World Object Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Open world object detection (OWOD) extends closed-set detection by identifying unknown objects and incrementally expanding the known class set. However, its assumption that training and evaluation data share the same domain limits applicability in real-world deployment, where visual domains continually change. Our experiments showed that even a state-of-the-art OWOD detector suffered performance degradation under such shifts, and a straightforward combination with a domain-generalized object detection (DGOD) method provided only marginal recovery. To bridge this gap, we introduce **Single-Domain Generalizable Open World Object Detection (SG-OWOD)**, where a detector trained on a single source domain must preserve its open world capability across multiple unseen target domains. To investigate these challenges, we analyzed a vision-language-based OWOD baseline. We found that domain shifts were associated with feature drift across the backbone and weaker known-class responses. Its attribute-selection statistics were also dominated by background, obscuring object-related evidence for unknown discovery. These observations motivate **RADAR** (**R**obust **A**ttribute-guided **D**etection via **A**ligned **R**epresentations), a framework built on a pretrained vision-language object detector (VLOD) with two complementary components. *Consistency-driven Low-rank Adaptation (CLA)* adapts early backbone layers by enforcing prediction consistency under style perturbations, improving known-class detection under domain shifts. *Background-Decoupled Attribute Selection (BDAS)* builds an object-centric attribute bank by separating object-like regions from background during selection, improving unknown recall when combined with CLA. Across three target domain benchmarks, RADAR improved over the strongest published baselines by at least K-mAP at Task 4 and H-Score at Task 3, while outperforming a naive OWOD and DGOD combination at every incremental task.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Open World Object Detection, Domain Generalization

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 5 reviews, above 27% of the others.
Reviewer_5Y1SOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

Introduced the Single-Domain Generalizable Open World Object Detection (SG-OWOD) setting, where a detector trained on a single labeled source domain must preserve its open-world capability across multiple unseen target domains and also perform class incremental learning. They performed analysis of vision language based OWOD baselines and revealed two major bottlenecks. First, severe feature drift across the backbone, accompanied by weaker known class responses under domain shifts, and background affected the unknown discovery. So, they proposed RADAR to maintain known class detection across domains and removed effect of background to improve unknown class detection to operate reliably across unseen domains.

Critical strengths
  • Tackles an important problem setting SG-OWOD by bridging the gap between open-world object detection and domain generalization to handle real-world distribution shifts using training data from only a single source domain.
  • Lightweight adaptation to address feature drift problem using CLA.
  • Filters out background clutter to improve unknown class detection.
  • Comprehensive empirical analysis over 3 types of shifts, to show improvement on distribution shifts and class incremental learning.
Critical weaknesses
  • CLA doesnt improve robustness as it improves the clean source domain about as much as the target domain. At source Task 4, K-mAP goes from 47.0 to 59.6, and previous class AP from 45.3 to 61.4. Relative robustness metrics could help show the improvement on target domain better.
  • It lacks on the evaluation of real world datasets like OdinW, flickr etc, also LVIS would be better to evaluate for unknown class detection. Since YOLO-World is pretrained on data that includes coco unknown classes content, as object365 consists of many overlapping coco classes. -The OVOW+DivAlign on COCO-DC is drastically worse than OVOW alone, which is unexplained. Also there are several other baselines that are missing, like simple strong augmentation AugMix on OW-OVD, full or partial fine-tuning without lora, and other VLM-based OWOD methodsthat the paper cites, such as OW-VAP and YOLO-UniOW.
  • findings of earlier layer suffering from distribution shifts and background affecting the unknown detections are not novel, and proposed method to improve robustness is not clearly proved.
Questions for authors
  • Could you evaluate the relative robustness metrics for all the methods and component ablations?
  • How does RADAR perform on benchmarks with natural domain shift, like ODinW, and on an unknown-class evaluation with categories less covered by pretraining, likein the LVIS rare split? -Several baselines are missing like OW-OVD with strong augmentation such as AugMix, OW-OVD with full or partial fine-tuning without LoRA, OW-OVD+DivAlign and the VLM-based OWOD methods like OW-VAP and YOLO-UniOW.
Reviewer_ug2DOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Content Summary

  • This paper introduces Single-domain Generalizable Open World Object Detection (SG-OWOD) setting, where a detector trained on a single source domain must perform both unknown object discovery and class-incremental learning across unseen target domains. To address this, the authors propose RADAR, a framework built upon YOLO-World-L that combines (1) Consistency-driven Low-rank Adaptation (CLA), a task-incremental LoRA tuning method on early backbone layers using style perturbations, and (2) Background-Decoupled Attribute Selection (BDAS), which separates negative anchors into object-like and background sets for attribute selection.

Contributions

  • Authors formulate Single-Domain Generalizable Open World Object Detection(SG-OWOD), coupling unknown object discovery and class-incremental learning under unseen domain shifts from a single source.
  • Authors propose Consistency-driven Low-rank Adaptation(CLA) which adapts early backbone layers via LoRA while enforcing prediction consistency across style-perturbed views to mitigate backbone feature drift.
  • Authors propose Background-Decoupled Attribute Selection(BDAS) which decouples negative anchors into object-like and background pools via JSD contrast to construct an object-centric attribute bank for unknown discovery.

Validation

  • The proposed model is trained incrementally over four tasks under the S-OWODB split (19/21/20/20 classes) using only clean COCO as the single source domain. All hyperparameter selection and development controls are conducted on source-domain validation data without access to target images. The frozen detector is then evaluated zero-shot across unseen target benchmarks—COCO-C (15 corruptions), COCO-O (6 style shifts), COCO-DC (4 background shifts), and Cityscapes →\rightarrow Foggy Cityscapes—reporting domain-averaged Known mAP (K-mAP), Unknown Recall (U-Recall), and their harmonic mean (H-Score).
Critical strengths

Practical and Clear Problem Formulation

  • The paper identifies a realistic limitation of existing OWOD methods—the assumption that training and deployment share the same visual domain—and clearly formulates the Single-Domain Generalizable OWOD (SG-OWOD) setting. Showing that a straightforward combination of a recent OWOD detector and a Single-DGOD method (OVOW + DivAlign) yields only marginal recovery effectively motivates the problem.

Parameter- and Inference-Efficient Adaptation

  • By restricting adaptation to low-rank residuals (LoRA) in the early backbone layers and current-task text embeddings, RADAR updates only ~0.4% of the detector's parameters (191K LoRA parameters) for a single source epoch per task. Moreover, because the learned LoRA residuals can be folded into the convolutional weights after training, it preserves the base detector's architecture at inference time.

Broad Evaluation Across Diverse Domain Shifts

  • The experimental evaluation covers a wide range of distribution shifts, including 15 synthetic corruptions (COCO-C), 6 natural style shifts (COCO-O), 4 background-composition shifts (COCO-DC), and a driving weather shift (Cityscapes →\rightarrow Foggy Cityscapes), alongside detailed per-domain breakdowns in the Appendix.
Critical weaknesses

Difficulty in Interpreting Key Observations and Limited Novelty in Problem Diagnosis The observation that domain shifts are associated with feature drift in early backbone layers (which encode style statistics) is already a well-established fact in the domain generalization literature (e.g., AdaIN, MixStyle) rather than a novel finding. Furthermore, the key observations motivating the method are relegated to Appendix C with insufficient explanation in the main text. In particular, Figure 2(b) is very difficult to interpret from the figure and caption alone (e.g., what distributions are being compared via JSD and what the axes represent), making it hard to clearly assess the contribution.

Questionable Necessity of BDAS and Severe Degradation in Unknown Precision

  • The main paper only reports Unknown Recall (U-R) and omits precision-related metrics for unknown objects. Frankly, if a model is encouraged to assign high objectness scores to any region, recall can easily be inflated. In a controlled incremental evaluation on public datasets, precision/AP over unknown objects should be evaluated.
  • Looking at Appendix Tables 22 and 23, unknown precision (U-AP) actually collapses when adding BDAS at Tasks 2 and 3 (e.g., dropping from 13.2 with +CLA alone to 6.7 with +BDAS+CLA on COCO-C Task 2, and from 15.9 to 7.2 on Source). Given this severe drop in U-AP, it is difficult to see why BDAS is beneficial.

Ambiguity and Feature-Space Mismatch in BDAS

  • Unclear Partitioning of Negative Anchors: While BDAS splits negative anchors into an object-like set (O\mathcal{O}) and a background set (B\mathcal{B}), even after consulting the Appendix (Algorithm 1), the process remains unclear.
  • Mismatch Between F0F_0 and FtF_t : Why are activation distributions always extracted using only the initial frozen encoder F0F_0 rather than F1,F2,…,Ft−1F_1, F_2, \dots, F_{t-1}? As class-incremental adaptation proceeds, the feature space of FtF_t drifts away from F0F_0, which could make the F0F_0-selected attribute bank suboptimal for the adapted detector.
  • Lack of Justification for the Scoring Formula : In Eq. 9, BDAS linearly combines JSD and mean cosine similarity via β\beta. The rationale for calibrating these two heterogeneous quantities—and whether the mean cosine similarity term is genuinely necessary—is not ablated or justified.

Lack of Analysis on Potential Failure Modes of CLA and Missing Sub-Component Ablations

  • Risk of Aggressive Style Extrapolation: CLA perturbs channel-wise feature statistics via extrapolation with a large factor (ϵ=5\epsilon = 5). Such aggressive extrapolation can push features out-of-manifold and produce non-sensical representations, yet there is no analysis verifying the validity of these perturbed features.
  • Unchanged Drift Trend & Missing LoRA-only Baseline: Appendix Table 10 shows that the increasing drift trend toward deeper layers remains unchanged after CLA, while stem drift even increases across all benchmarks. To justify CLA's design, an ablation comparing vanilla LoRA (without style perturbation) against the full CLA (with the style bank perturbation and consistency loss) is essential, as vanilla LoRA is already a standard approach.
Questions for authors
  • Mitigating Catastrophic Forgetting: Catastrophic forgetting is inherent to this incremental setup, yet the paper relies almost entirely on a naive EMA-like weight fusion with a small merge factor (τt>1=0.1\tau_{t>1}=0.1, where Table 16 shows τ2=1.0\tau_2=1.0 collapses on previous classes). Have the authors explored more principled continual learning mechanisms, and how sensitive is the proposed method to different task orderings ?

  • Altering the Feature Drift Trend: While CLA lowers the absolute drift in later stages, the upward drift trend across backbone depth persists. Have the authors considered ways to fundamentally alter this trend?

  • Target Deployment Scenario & Few-Shot Inference: Since visual features are adapted in the image encoder while class information is injected via text embeddings, could the authors clarify the exact target deployment scenario, and whether this setup supports few-shot inference ?

  • Clarifications on Baseline Combination and "CoRe": Appendix F (Tables 25 and 26) mentions "closed-form re-anchoring (CoRe)", which never appears in the main paper. Could the authors clarify what CoRe is?

Reviewer_nGRuOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Single-Domain Generalizable Open World Object Detection (SG-OWOD), which combines unknown object discovery, class-incremental learning, and generalization to unseen domains. The proposed RADAR framework builds upon OW-OVD and consists of two components: Consistency-driven Low-rank Adaptation (CLA) for improving cross-domain robustness and Background-Decoupled Attribute Selection (BDAS) for enhancing unknown detection. Experiments on COCO-C, COCO-O, COCO-DC, and Foggy Cityscapes demonstrate improvements over existing baselines.

Critical strengths
  1. Meaningful problem setting. SG-OWOD extends existing open-world detection to unseen domains, addressing a practically relevant challenge that has received limited attention.

  2. Comprehensive experiments. The paper evaluates multiple domain shifts and incremental learning stages, demonstrating competitive performance improvements with supporting ablation studies.

Critical weaknesses
  1. Limited methodological novelty and insufficient modeling of coupled challenges. RADAR primarily addresses domain generalization and unknown discovery through separate components, without explicitly modeling their interactions. Moreover, CLA largely combines established techniques, including style perturbation, prediction consistency, LoRA, and incremental weight merging. The technical novelty beyond integrating existing approaches remains unclear.

  2. Questionable improvements in unknown detection quality. Table 18 shows that unknown-score scaling substantially increases U-Recall (38.4 to 67.1 on COCO-C), while U-AP improves only marginally (15.6 to 16.7). Furthermore, Table 23 shows that BDAS consistently reduces U-AP at Task 3 across all three benchmarks. These findings raise concerns about whether the improvements reflect better unknown object discovery rather than score adjustment.

Questions for authors
  1. What unique challenges arise from the interaction between domain generalization and open-world incremental learning, and how does RADAR explicitly address these interactions rather than treating them independently?

  2. What are the key technical contributions of CLA beyond existing style perturbation, consistency learning, and LoRA-based adaptation? Could the authors provide separate ablations for its consistency objectives?

  3. Given the strong influence of unknown-score scaling and the U-AP degradation in Table 23, can the authors further demonstrate that BDAS genuinely improves unknown detection quality?

  4. Since RADAR builds upon pretrained YOLO-World-L, are the evaluated unknown categories genuinely unseen during pretraining, or only excluded from the current incremental learning stage?

Reviewer_qRW1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies Single-Domain Generalizable Open-World Object Detection (SG-OWOD), combining cross-domain generalization, unknown-object detection, and incremental learning. The proposed RADAR framework builds on OW-OVD and YOLO-World and introduces two components: Consistency-driven Low-rank Adaptation (CLA) for domain generalization and Background-Decoupled Attribute Selection (BDAS) for unknown-object detection. The method is evaluated on COCO-C, COCO-O, and COCO-DC across four incremental tasks.

Critical strengths
  1. Meaningful problem setting and strong cross-domain detection results.

SG-OWOD addresses a challenging and practically relevant combination of detection problems. CLA provides substantial improvements in known-class detection across unseen target domains, particularly in later incremental tasks. For example, Table 3 shows that CLA improves COCO-C Task 4 known-class mAP from 29.6 to 39.6. Similar gains on COCO-O and COCO-DC support its empirical effectiveness.

  1. Reasonable motivation and extensive component analysis.

BDAS addresses a plausible limitation of OW-OVD: background-dominated unmatched proposals may bias attribute selection. Separating object-like regions from background is a reasonable approach, and BDAS improves unknown recall, especially at Task 1. The paper also provides useful component ablations and analyses of proposal mining, attribute selection, and unknown scoring.

Critical weaknesses
  1. The independent methodological contribution of CLA is insufficiently established.

CLA combines established techniques, including style perturbation, prediction consistency, and LoRA. Related work such as DivAlign has already explored view diversification and classification/localization consistency for domain-generalized detection.

Eqs. (7)–(8) introduce consistency terms for classification, bounding-box localization, and attribute responses. Attribute consistency appears to be a distinctive component for SG-OWOD, but the current ablations do not isolate its contribution from the existing techniques.

Although the combined CLA module is effective, it remains unclear whether its proposed design provides a significant methodological improvement beyond established domain-generalization approaches.

  1. BDAS does not consistently improve unknown-detection quality.

Appendix Tables 22–23 show that adding BDAS to CLA reduces unknown AP at Tasks 2 and 3 across all three target benchmarks.

For example, on COCO-C, adding BDAS changes:

Task 2: U-Recall 63.4 → 67.1, but U-AP 13.2 → 6.7. Task 3: U-Recall 62.7 → 68.0, but U-AP 8.4 → 5.2. These results indicate a substantial precision–recall trade-off. Since H-Score combines known-class mAP with unknown recall but excludes unknown AP, improvements in H-Score alone do not establish better overall unknown-detection quality.

This limitation is significant because BDAS is one of the paper’s two main contributions.

Questions for authors
  1. What is the independent contribution of attribute consistency within CLA?

Please compare the following configurations under the same backbone, LoRA settings, training budget, and inference scoring procedure:

A: LoRA with original-view detection loss.

B: A + style-perturbed detection loss.

C: B + classification and box consistency.

D: C + attribute consistency (full CLA).

Report known-class mAP and unknown AP on both source and target domains. In particular, the D versus C comparison would directly quantify the added value of attribute consistency.

  1. Why does BDAS substantially reduce unknown AP in later tasks?

Please provide unknown precision–recall curves or recall at matched false-positive rates for CLA versus CLA+BDAS at Tasks 2 and 3.

Additionally, using evaluation annotations, compare the fraction of mined object-like proposals corresponding to actual unknown objects versus background across Tasks 1–3. This could help determine whether changes in proposal-mining quality contribute to the AP reductions.

  1. Does CLA specifically improve domain robustness, beyond improving detection accuracy in general?

In Table 3, CLA improves Task 4 source known-class mAP from 47.0 to 59.6 and COCO-C target mAP from 29.6 to 39.6.

Please compare the relative source-to-target performance drop for Base and +CLA, using matched source and target evaluations. This would help distinguish improved domain robustness from general detection improvements.

Additional Comments

  1. Unknown-score scaling: Eq. (11) uses ξ_u = 15. Appendix Table 18 shows that on COCO-C Task 1, scaling increases U-Recall from 38.4 to 67.1, while U-AP changes only from 15.6 to 16.7. Please clarify the role of score scaling when interpreting U-Recall and H-Score, and summarize the later-task U-AP trade-off in the main paper.

  2. Inherited components: Please distinguish the new contributions of CLA and BDAS from the attribute scoring, vision-language representations, and incremental-learning procedures inherited from OW-OVD.

Overall Assessment RADAR addresses an important problem and demonstrates substantial cross-domain known-class detection gains. However, the independent novelty of CLA remains unclear, while BDAS produces significant unknown-AP reductions in later tasks.

The key question is whether CLA alone provides a sufficiently distinct and significant contribution to justify acceptance. More focused ablations and an explanation of the BDAS precision–recall trade-off could change my assessment.

Reviewer_pAL9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Single-Domain Generalizable Open World Object Detection (SG-OWOD), where a detector learns classes incrementally from a single source domain and detects known and unknown objects in unseen target domains. The proposed framework, RADAR, combines Consistency-driven Low-rank Adaptation (CLA) to improve robustness under domain shifts with Background-Decoupled Attribute Selection (BDAS) to select object-centric attributes for unknown object discovery.

Critical strengths

1.The paper addresses a practically relevant problem by jointly considering domain generalization, unknown object discovery, and class-incremental learning. The proposed setting provides a useful perspective on the reliability of open world detectors beyond their training domain.

2.The evaluation covers corruption, appearance, and background shifts on COCO-C, COCO-O, and COCO-DC, respectively, as well as weather shifts through the Cityscapes → Foggy Cityscapes experiment.

Critical weaknesses

1.The methodological novelty of CLA is not sufficiently established in the main text. CLA combines low-rank adaptation, style perturbation, and prediction consistency, all of which have established precedents acknowledged by the paper. Although their integration is reasonable, the paper does not clearly establish which design choices address challenges specific to SG-OWOD or why a simpler combination would be insufficient.

2.A substantial portion of the reported improvement comes from changing the unknown scoring function. In Table 3, replacing OW-OVD’s scoring function alone increases the COCO-C Task 1 H-Score from 32.3 to 55.0, whereas the full model achieves 59.2. The common-score component ablations are useful, but the main comparisons do not establish whether RADAR retains its advantage when competing methods receive equivalent source-only score calibration.

3.The main-text ablations evaluate CLA as a whole without isolating style perturbation, the individual consistency losses, or task-dependent update scaling. Consequently, the necessity of these choices remains unclear. This also limits the evidence connecting the observed improvements to the proposed mechanism of mitigating domain-induced feature drift.

Questions for authors

1.How much of RADAR’s advantage remains when strong baselines receive equivalent source-only calibration of their unknown scores, using the same calibration criterion and evaluation protocol?

2.Could the authors provide controlled ablations of style perturbation, classification consistency, box consistency, attribute consistency, and task-dependent update scaling to establish which components are necessary for the reported improvements?

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

Related papers

Discussion (0)

Sign in to comment.