Pith. sign in

REVIEW 4 major objections 4 minor 60 references

Open-vocabulary object detection fails under domain shift because visual features drift away from their text-defined anchors, and a curriculum that trains on reliable regions first can keep the alignment space stable.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 17:10 UTC pith:X6MMFJWF

load-bearing objection A useful DG-OVOD benchmark and a plausible curriculum method, but the headline gains are modest and confounded by test-set tuning and a self-referential difficulty estimator; worth a serious referee, not yet a robust result. the 4 major comments →

arxiv 2603.27556 v2 pith:X6MMFJWF submitted 2026-03-29 cs.CV

Towards Domain-Generalized Open-Vocabulary Object Detection: A Progressive Domain-invariant Cross-modal Alignment Method

Xiaoran Xu , Xiaoshan Yang , Jiangang Yang , Yifan Xu , Jian Liu , Changsheng Xu This is my paper
classification cs.CV
keywords domain generalizationopen-vocabulary object detectioncross-modal alignmentcurriculum learningrobustness to distribution shiftpseudo-word prototypesalignment drift
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Open-vocabulary object detection (OVOD) promises to recognize any category at inference, but its reliability under visual domain shifts remains an open problem. This paper argues that the core vulnerability is not image noise per se, but a destabilization of the latent space where visual features are matched to text-defined category anchors: novel categories, which lack direct supervision, drift away from their semantic anchors, causing a collapse in cross-modal alignment. The paper formalizes this as Domain-Generalized OVOD (DG-OVOD), proposes two robustness benchmarks, and shows that current OVOD methods lose on average more than half of their novel-class accuracy under corruption and style shift. To address this, it introduces PICA, a curriculum that ranks image regions by two complementary proxies—ambiguity (how close a region's feature is to a wrong category) and signal strength (how confidently it matches its own category)—and trains the cross-modal projection first on low-ambiguity, high-confidence regions, gradually adding harder ones. In the paper's experiments, PICA improves average novel-class mAP from 19.8 to 20.7 on the corruption benchmark and from 18.5 to 19.7 on the out-of-distribution benchmark. The central takeaway is that OVOD robustness is governed by the stability of the cross-modal alignment space, and that a training curriculum can enforce that stability.

Core claim

The paper's central claim is that the stability of the latent cross-modal alignment space is the binding constraint on open-vocabulary detection under domain shift. Through empirical analysis, it shows that visual shifts cause novel-category visual signals to detach from their text-derived semantic anchors, and that this destabilization is sample-dependent, with some regions suffering signal collapse (lower positive alignment) and others boundary confusion (shrinking margin to negatives). The proposed method, PICA, enforces domain-invariant alignment by a progressive curriculum: regions are scored by ambiguity h and signal strength q, partitioned into difficulty tiers, and sampled in increas

What carries the argument

The mechanism is a curriculum built from two per-region proxies in the text-embedding space: signal strength q = sim(f_r, w_r), the cosine similarity between a region's visual feature and its own pseudo-word prototype, and ambiguity h = s^- - s^+, the margin between the hardest negative prototype and the positive. Pseudo-word prototypes w_r = P_img_to_text(f_r) are produced by a learnable projection from visual features, which is grounded by the open-vocabulary loss. Regions are split into easy/medium/hard tiers by h, a soft quality gate penalizes the bottom percentile of q within each tier, and a scheduler alpha(rho) progressively raises the sampling ratio of hard tiers while a bidirectiona

Load-bearing premise

The curriculum's hardness scores are computed using the same learnable projection that PICA is training, so if the projection is itself corrupted by the domain shift, the sample selection may be based on an unreliable self-assessment rather than true cross-modal confusion.

What would settle it

Train PICA on OV-COCO-C but measure h and q with a frozen, domain-robust projection (e.g., a pretrained text-aligned visual encoder). If the ranking of easy vs hard regions changes substantially from the learned projection, or if the frozen-projection curriculum outperforms the learned-projection one, the paper's assumption that the learned projection is trustworthy for sample selection is refuted. Alternatively, an experiment showing that PICA's performance improvement vanishes when the AI-gap is held constant would weaken the causal link.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • The proposed DG-OVOD benchmark provides a common protocol for evaluating open-vocabulary detection under both synthetic corruptions and natural out-of-distribution domains.
  • PICA improves novel-category detection under corruption from 19.8 to 20.7 mAP50 and under natural OOD from 18.5 to 19.7, while keeping clean-domain accuracy roughly constant.
  • The ablation shows that both the ambiguity and signal-strength proxies are necessary: using either alone underperforms the combination, and removing the curriculum degrades robustness.
  • The alignment-invariance gap analysis indicates that PICA's gains coincide with a smaller gap between clean and corrupted logit vectors, supporting the paper's hypothesis that stability of the cross-modal space is the key factor.
  • The gradient-direction cosine similarity shows that PICA maintains smoother, more consistent optimization than baseline, implying fewer gradient conflicts from unreliable regions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the proxies are computed by the same projection being trained, the curriculum's benefit may shrink for extreme stylizations that corrupt the projection itself—the paper's flat or negative results on Tattoo and Sketch are consistent with this ceiling.
  • A natural extension would be to compute h and q using an ensemble or a frozen, robust text encoder, which could decouple the curriculum signal from the learned projection and potentially extend gains to extreme domains.
  • The same curriculum principle could be applied to other cross-modal alignment tasks (e.g., open-vocabulary segmentation, grounding, or retrieval), with the proxies adapted to their region or token granularity.
  • The AI-gap could be turned from a diagnostic into an explicit regularizer, penalizing mismatch between clean and corrupted logits during training; the paper stops short of that, but its data suggest it would help.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper formalizes Domain-Generalized Open-Vocabulary Object Detection (DG-OVOD) and proposes PICA, a curriculum-based training method that sorts region samples by an ambiguity proxy h and a signal-strength proxy q, computed from pseudo-word prototypes produced by the trainable projection P_img_to_text. PICA progressively increases the proportion of hard samples during training, with a soft quality gate that downweights signal-collapsed regions. The authors introduce two benchmarks (OV-COCO-C and OV-COCO-O) and report improvements over existing OVOD methods, arguing that robustness under domain shift is intrinsically linked to the stability of the latent cross-modal alignment space. The empirical section includes ablation studies, an alignment-invariance gap analysis, gradient-direction similarity, and a hyperparameter sensitivity study.

Significance. The paper contributes a new evaluation protocol for a relevant problem (OVOD under distribution shift) and a detailed empirical analysis of cross-modal alignment degradation. The proposed curriculum is well-motivated and the ablations are systematic. However, the central claim is not adequately supported: the difficulty proxies are self-referential to the very projection being trained, the hyperparameters are tuned on the test benchmark, and no uncertainty quantification is provided. The method also underperforms on the extreme domains (Tattoo, Sketch) where the paper admits the proxies fail. If the validation issues are fixed, the benchmark and methodology could be useful to the community, but in the current form the evidence is not convincing.

major comments (4)
  1. [Sec. 3.2, Eqs. (3)–(5); Sec. 4.2, Table 2] The difficulty proxies h and q are computed from pseudo-word prototypes w_r = P_img_to_text(f_r), where P_img_to_text is the same projection that PICA trains. The paper itself concedes in Table 2 that on Tattoo and Sketch 'signal collapse undermines the reliability of our cross-modal alignment proxies q and h'. This directly undermines the central claim: the curriculum fails exactly under the shifts it is designed to address. The authors should provide evidence that h/q faithfully measure semantic alignment, e.g., by correlating them with an oracle measure using true class anchors, or by showing that the source-domain curriculum selection predicts target-domain robustness across domains.
  2. [Sec. 4.3, Table 5] Hyperparameters such as base curriculum ratios (r0_E, r0_M, r0_H), queue size M, and presumably p_q, delta, and alpha are selected by evaluating on the OV-COCO-C test benchmark itself. Table 5 sweeps (r0_E, r0_M, r0_H) and M and reports the best mAPavg, which is then used as the final setting. This is post-hoc selection on the test set and invalidates the reported comparison against baselines that were not tuned in the same way. The authors should use a separate validation split (e.g., a subset of corruption types) or fix defaults a priori.
  3. [All experimental tables] Every result is reported as a single run with no error bars, multiple seeds, or significance tests. The gains are modest (20.7 vs 19.8 on OV-COCO-C; 19.7 vs 18.5 on OV-COCO-O), and without uncertainty estimates it is impossible to assess whether the improvements are statistically meaningful. The paper should report mean and standard deviation over at least three seeds, and ideally a paired test against the main baseline.
  4. [Algorithm 1] Algorithm 1 is internally inconsistent regarding the ordering of difficulty tiers. Equation (5) defines larger h as more ambiguous, so sorting by h ascending puts the least ambiguous samples in the first tier. However, line 2 labels Tier 1 as 'Hardest' and Tier K as 'Easiest', which is the reverse. The curriculum schedule in Step 4 increases the 'Hard' ratio over time, so the mapping between tiers, h, and the ratios must be clarified. As written, the algorithm is not reproducible.
minor comments (4)
  1. [Table 1 caption] The sentence 'R.CLIP, Obj and DQUO refers to [58] [1] [48]' is ungrammatical and should be 'refer to'. Also, the definitions of Baron* and Baron† should be stated more clearly in the caption, not only in the text.
  2. [Eq. (10)] The symbol sim is overloaded: it is used both for cosine similarity between vectors and for the similarity between two logit vectors. Please use different notation, e.g., CosSim for the former and a separate name for the latter.
  3. [Sec. 4.2] The comparison on OV-COCO-C uses Baron† (Baron + Mixup + DeepCORAL) as the strongest baseline, but the table also shows Baron and two other variants. The text should explicitly state which baseline is used for the claimed improvement, and why this baseline is the most appropriate.
  4. [Sec. 4.4] The AI-gap analysis (Fig. 3a) shows the proposed method has lower gaps, but it is not accompanied by a statistical or correlation analysis linking AI-gap to mAP across methods or domains. Adding such a correlation would strengthen the central claim that alignment stability is intrinsically linked to robustness.

Circularity Check

0 steps flagged

No significant circularity: PICA's evaluation is externally benchmarked, and its self-referential difficulty proxies are a standard curriculum-learning design, not a circular derivation.

full rationale

I examined the derivation chain: the pseudo-word prototypes w_r = P_img_to_text(f_r) (Eq. 1), the signal-strength and ambiguity proxies q and h (Eqs. 2-5), the curriculum sampler (Algorithm 1), the curriculum loss (Eq. 6), and the final benchmark results (Tables 1-2). The difficulty proxies are indeed computed using the same trainable projection P_img_to_text that PICA trains, which makes sample selection self-referential in the sense common to self-paced learning. However, this is not a circular derivation of the paper's central empirical claim: PICA's reported improvements are measured on held-out OV-COCO-C and OV-COCO-O benchmarks against external baselines, and the pseudo-word prototypes are explicitly grounded to text embeddings by the external open-vocabulary loss L_ovod (stated in Section 3.1). The paper also honestly concedes a limitation in Table 2: 'extreme stylization and sparse structure cause signal collapse that undermines the reliability of our cross-modal alignment proxies q and h.' This is a stated weakness of the proxy under certain shifts, not an admission that the benchmark result is equivalent to the proxy by construction. No load-bearing self-citation, no imported uniqueness theorem, and no ansatz smuggled in via citation were found. The 'intrinsic link' between alignment stability and robustness is an empirical claim supported by external mAP evaluations and the AI-gap analysis, not a tautology. Thus the paper is self-contained against external benchmarks, and the correct finding is no significant circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The method's central load-bearing components are the semantic grounding of pseudo-word prototypes and the reliability of self-derived difficulty proxies; neither is independently verified. Hyperparameters are selected on the test benchmark.

free parameters (7)
  • base curriculum ratios (r0_E, r0_M, r0_H) = (0.33, 0.33, 0.33)
    Selected on OV-COCO-C in Table 5; balanced setting yields best mAP.
  • queue size M = 4096
    Selected on OV-COCO-C in Table 5; ablation shows 4096 best.
  • quality percentile p_q = 0.05
    Default chosen in Algorithm 1; no sensitivity analysis.
  • penalty delta = 2.0
    Default chosen in Algorithm 1; no sensitivity analysis.
  • scheduling function alpha(rho) = (2/3)*rho
    Default in Algorithm 1; not ablated.
  • curriculum loss weight lambda_curr = 1.0
    Chosen in Section 4.1; no sensitivity analysis.
  • number of tiers K = 3
    Default in Algorithm 1; fixed.
axioms (5)
  • domain assumption CLIP text/visual encoders provide a shared semantic space stable enough for OVOD under domain shift.
    Section 3.1 assumes F_text and F_vision from a pretrained VLM are fixed and usable, and that novel-class text embeddings remain valid anchors under visual shift.
  • domain assumption Pseudo-word prototypes w_r = W f_r + b are grounded in the language embedding space by L_ovod, so negative samples Z^- encode semantic supervision.
    Section 3.2 states this to justify h/q as genuine cross-modal confusion; no independent verification is provided.
  • ad hoc to paper Signal collapse and boundary confusion are orthogonal failure modes fully captured by q_r and h_r.
    Proposed in Section 3.2; used to define the curriculum tiers.
  • domain assumption Bidirectional InfoNCE over sampled regions prevents latent collapse and enforces visual consistency.
    Section 3.2, Eq (6); this is an empirical claim about the training objective.
  • domain assumption mAP50 aggregated hierarchically over corruptions/domains is a fair protocol for DG-OVOD.
    Section 4.1, Eqs (8)-(9); choice of evaluation metric.

pith-pipeline@v1.3.0-alltime-deepseek · 13868 in / 11762 out tokens · 113395 ms · 2026-08-02T17:10:01.830880+00:00 · methodology

0 comments
read the original abstract

Open-Vocabulary Object Detection (OVOD) has achieved remarkable success in generalizing to novel categories. However, this success often rests on the implicit assumption of domain stationarity. In this work, we revisit the OVOD paradigm and study a key vulnerability: the fragile coupling between visual manifolds and textual embeddings under distribution shifts. We first formulate Domain-Generalized Open-Vocabulary Object Detection (DG-OVOD) as an evaluation protocol for open-vocabulary recognition under visual shifts. Through empirical analysis, we observe that visual shifts can destabilize the latent cross-modal space, causing novel-category visual signals to drift away from their semantic anchors. Motivated by these observations, we propose Progressive Domain-invariant Cross-modal Alignment (PICA). PICA departs from uniform training by introducing a multi-level curriculum based on ambiguity and signal strength. It constructs a quality-adjusted curriculum over pseudo-word prototypes, refined by sample reliability and visual consistency, to encourage more stable cross-domain modality alignment. Our findings suggest that OVOD robustness under domain shifts is closely linked to the stability of the latent cross-modal alignment space. Our work provides a DG-OVOD evaluation protocol and a practical perspective on building more generalizable open-vocabulary systems beyond static laboratory conditions.

Figures

Figures reproduced from arXiv: 2603.27556 by Changsheng Xu, Jiangang Yang, Jian Liu, Xiaoran Xu, Xiaoshan Yang, Yifan Xu.

Figure 1
Figure 1. Figure 1: Cross-modal alignment analysis. (a) PICA preserves more stable and object￾centric attention patterns across both base and novel categories compared to Baron [52]. (b) The difficulty of preserving cross-modal alignment under domain shifts varies across different samples. paradigm that transcends traditional closed-set detection [7, 13, 32, 55, 58, 59]. By projecting natural language category descriptions in… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the Progressive Domain-invariant Cross-modal Alignment (PICA). The progressive sampler ranks visual and textual region features by ambiguity proxy h and signal strength proxy q, dividing them into tiers. It then uses a dynamic sampling ratio α(ρ), where ρ represents the training iteration, to align these features with pseudo￾word prototypes in a staged curriculum to enhance cross-modal consiste… view at source ↗
Figure 3
Figure 3. Figure 3: Comparative analysis of cross-modal alignment and training stability. (a) Mean AI-gap across OOD domains. (b) Gradient cosine similarity during training. (a) Baseline exhibits persistent high-∆h dis￾persion across mid-to-high signal levels. The q decreases as training progresses. (b) PICA progressively stabilizes mid-to-high signal regions (∆h → 0). And the q maintains stability as training goes on [PITH_… view at source ↗
Figure 4
Figure 4. Figure 4: Region-level cross-modal alignment stability between standard and Gaussian noise. Each point represents a region, plotted by its clean-image signal strength q (x￾axis) and ambiguity h (y-axis), colored by confusion increase ∆h = hcorrupted − hclean, demonstrating that curriculum training selectively consolidates alignment robustness for reliable regions. 4.4 Analysis To further examine how PICA improves cr… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

60 extracted references · 14 linked inside Pith

  1. [1]

    Advances in Neural Information Processing Systems35, 33781–33794 (2022) 2, 4, 6, 10, 11

    Bangalath, H., Maaz, M., Khattak, M.U., Khan, S.H., Shahbaz Khan, F.: Bridging the gap between object and image-level representations for open-vocabulary detec- tion. Advances in Neural Information Processing Systems35, 33781–33794 (2022) 2, 4, 6, 10, 11

    2022

  2. [2]

    In: Pro- ceedings of the 26th annual international conference on machine learning

    Bengio, Y., Louradour, J., Collobert, R., Weston, J.: Curriculum learning. In: Pro- ceedings of the 26th annual international conference on machine learning. pp. 41–48 (2009) 5

    2009

  3. [3]

    arXiv preprint arXiv:1805.04807 (2018) 5

    Cai, Q.Z., Du, M., Liu, C., Song, D.: Curriculum adversarial training. arXiv preprint arXiv:1805.04807 (2018) 5

    Pith/arXiv arXiv 2018

  4. [4]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Chen, C., Xie, W., Huang, W., Rong, Y., Ding, X., Huang, Y., Xu, T., Huang, J.: Progressive feature alignment for unsupervised domain adaptation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 627– 636 (2019) 4

    2019

  5. [5]

    Advances in Neural Information Processing Systems36, 51758–51777 (2023) 2

    Chen, S., Gu, J., Han, Z., Ma, Y., Torr, P., Tresp, V.: Benchmarking robustness of adaptation methods on pre-trained vision-language models. Advances in Neural Information Processing Systems36, 51758–51777 (2023) 2

    2023

  6. [6]

    In: European conference on computer vision

    Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: Uniter: Universal image-text representation learning. In: European conference on computer vision. pp. 104–120. Springer (2020) 1

    2020

  7. [7]

    arXiv preprint arXiv:2401.17270 (2024) 2

    Cheng, T., Huang, Y., Ding, M., Wang, Y., Zhu, Y., Li, G., , et al.: Yolo-world: Real-time open-vocabulary object detection. arXiv preprint arXiv:2401.17270 (2024) 2

    Pith/arXiv arXiv 2024

  8. [8]

    In: European Confer- ence on Computer Vision

    Chhipa, P.C., De, K., Chippa, M.S., Saini, R., Liwicki, M.: Open-vocabulary object detectors: Robustness challenges under distribution shifts. In: European Confer- ence on Computer Vision. pp. 62–79. Springer (2024) 2, 9

    2024

  9. [9]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Cubuk,E.D.,Zoph,B.,Mane,D.,Vasudevan,V.,Le,Q.V.:Autoaugment:Learning augmentation strategies from data. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 113–123 (2019) 4

    2019

  10. [10]

    In: 2009 IEEE conference on computer vision and pattern recognition

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large- scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248–255. Ieee (2009) 4

    2009

  11. [11]

    Journal of machine learning research17(59), 1–35 (2016) 4

    Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., March, M., Lempitsky, V.: Domain-adversarial training of neural networks. Journal of machine learning research17(59), 1–35 (2016) 4

    2016

  12. [12]

    arXiv preprint arXiv:1710.02515 (2017) 5 16 X

    Graves, A., Bellemare, M., Menick, J., Munos, R., Osband, I.: Automated curricu- lum learning for neural networks. arXiv preprint arXiv:1710.02515 (2017) 5 16 X. Xu et al

    Pith/arXiv arXiv 2017

  13. [13]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Gu, M., Lin, G., Chen, S., Zhang, R., Wang, Y., Yang, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1112– 1123 (2021) 2, 3

    2021

  14. [14]

    arXiv preprint arXiv:2007.01434 (2020) 4

    Gulrajani, I., Lopez-Paz, D.: In search of lost domain generalization. arXiv preprint arXiv:2007.01434 (2020) 4

    Pith/arXiv arXiv 2007

  15. [15]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356–5364 (2019) 4

    2019

  16. [16]

    arXiv preprint arXiv:1903.12261 (2019) 4, 9

    Hendrycks,D.,Dietterich,T.:Benchmarkingneuralnetworkrobustnesstocommon corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019) 4, 9

    Pith/arXiv arXiv 1903

  17. [17]

    arXiv preprint arXiv:1912.02781 (2019) 4

    Hendrycks, D., Mu, N., Cubuk, E.D., Zoph, B., Gilmer, J., Lakshminarayanan, B.: Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781 (2019) 4

    Pith/arXiv arXiv 1912

  18. [18]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Huang, J., Guan, D., Xiao, A., Lu, S.: Fsdr: Frequency space domain randomiza- tion for domain generalization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6891–6902 (2021) 4

    2021

  19. [19]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2024) 4

    Huang, Z., Yang, M., Xiao, X., Hu, P., Peng, X.: Noise-robust vision-language pre- training with positive-negative learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024) 4

    2024

  20. [20]

    In: International conference on machine learning

    Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904–4916. PMLR (2021) 1, 5

    2021

  21. [21]

    arXiv preprint arXiv:1609.04836 (2016) 14

    Keskar, N.S., Mudigere, D., Nocedal, J., Smelyanskiy, M., Tang, P.T.P.: On large- batch training for deep learning: Generalization gap and sharp minima. arXiv preprint arXiv:1609.04836 (2016) 14

    Pith/arXiv arXiv 2016

  22. [22]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Kim, J., Cho, E., Kim, S., Kim, H.J.: Retrieval-augmented open-vocabulary object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17427–17436 (2024) 4, 6, 10

    2024

  23. [23]

    In: International conference on machine learning

    Kim, W., Son, B., Kim, I.: Vilt: Vision-and-language transformer without convo- lution or region supervision. In: International conference on machine learning. pp. 5583–5594. PMLR (2021) 1

    2021

  24. [24]

    In: Advances in neural information processing systems

    Kumar, M.P., Packer, B., Koller, D.: Self-paced learning for latent variable models. In: Advances in neural information processing systems. pp. 1189–1197 (2010) 5

    2010

  25. [25]

    In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition

    Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10965–10975 (2022) 4

    2022

  26. [26]

    In: European conference on computer vision

    Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740–755. Springer (2014) 4

    2014

  27. [27]

    Advances in Neural Information Processing Systems34, 18878–18890 (2021) 14

    Liu, B., Liu, X., Jin, X., Stone, P., Liu, Q.: Conflict-averse gradient descent for multi-task learning. Advances in Neural Information Processing Systems34, 18878–18890 (2021) 14

    2021

  28. [28]

    arXiv preprint arXiv:2108.13624 (2021) 4

    Liu, J., Shen, Z., He, Y., Zhang, X., Xu, R., Yu, H., Cui, P.: Towards out-of- distribution generalization: A survey. arXiv preprint arXiv:2108.13624 (2021) 4

    Pith/arXiv arXiv 2021

  29. [29]

    In: European conference on computer vision

    Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Jiang, Q., Li, C., Yang, J., Su, H., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In: European conference on computer vision. pp. 38–55. Springer (2024) 2 Abbreviated paper title 17

    2024

  30. [30]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Mao, X., Chen, Y., Zhu, Y., Chen, D., Su, H., Zhang, R., Xue, H.: Coco-o: A benchmark for object detectors under natural distribution shifts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6339–6350 (2023) 4, 9

    2023

  31. [31]

    arXiv preprint arXiv:1907.07484 (2019) 4

    Michaelis, C., Mitzkus, B., Geirhos, R., Rusak, E., Bringmann, O., Ecker, A.S., Bethge, M., Brendel, W.: Benchmarking robustness in object detection: Au- tonomous driving when winter is coming. arXiv preprint arXiv:1907.07484 (2019) 4

    Pith/arXiv arXiv 1907

  32. [32]

    In: European conference on computer vision

    Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Doso- vitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection. In: European conference on computer vision. pp. 728–755. Springer (2022) 2, 3, 5

    2022

  33. [33]

    Journal of Machine Learning Research21(181), 1–50 (2020) 5

    Narvekar, S., Peng, B., Leonetti, M., Sinapov, J., Taylor, M.E., Stone, P.: Curricu- lum learning for reinforcement learning domains: A framework and survey. Journal of Machine Learning Research21(181), 1–50 (2020) 5

    2020

  34. [34]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Pham, C., Vu, T., Nguyen, K.: Lp-ovod: Open-vocabulary object detection by lin- ear probing. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 779–788 (2024) 3

    2024

  35. [35]

    In: Proceedings of the IEEE international conference on computer vision

    Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazeb- nik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: Proceedings of the IEEE international conference on computer vision. pp. 2641–2649 (2015) 4

    2015

  36. [36]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PmLR (2021) 1, 2, 4, 5

    2021

  37. [37]

    arXiv preprint arXiv:2103.01946 (2021) 4

    Rebuffi, S.A., Gowal, S., Calian, D.A., Stimberg, F., Wiles, O., Mann, T.: Fixing data augmentation to improve adversarial robustness. arXiv preprint arXiv:2103.01946 (2021) 4

    Pith/arXiv arXiv 2021

  38. [38]

    Advances in neural information processing systems28(2015) 9, 10

    Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object de- tection with region proposal networks. Advances in neural information processing systems28(2015) 9, 10

    2015

  39. [39]

    In: European conference on computer vision

    Seo, S., Suh, Y., Kim, D., Kim, G., Han, J., Han, B.: Learning to optimize do- main specific normalization for domain generalization. In: European conference on computer vision. pp. 68–83. Springer (2020) 4

    2020

  40. [40]

    arXiv preprint arXiv:2403.16188 (2024) 2

    Shangguan, Z., Seita, D., Rostami, M.: Cross-domain multi-modal few-shot object detection via rich text. arXiv preprint arXiv:2403.16188 (2024) 2

    Pith/arXiv arXiv 2024

  41. [41]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8430–8439 (2019) 4

    2019

  42. [42]

    In: Proceed- ings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceed- ings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2556–2565 (2018) 4

    2018

  43. [43]

    In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion

    Singha, M., Jha, A., Bose, S., Nair, A., Abdar, M., Banerjee, B.: Unknown prompt the only lacuna: Unveiling clip’s potential for open domain generalization. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. pp. 13309–13319 (2024) 2

    2024

  44. [44]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Srinivasan,T.,Ren,X.,Thomason,J.:Curriculumlearningfordata-efficientvision- language alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5619–5624 (2023) 5 18 X. Xu et al

    2023

  45. [45]

    In: European conference on computer vision

    Sun,B.,Saenko,K.:Deepcoral:Correlationalignmentfordeepdomainadaptation. In: European conference on computer vision. pp. 443–450. Springer (2016) 10, 11

    2016

  46. [46]

    arXiv preprint arXiv:1412.3474 (2014) 10, 11

    Tzeng, E., Hoffman, J., Zhang, N., Saenko, K., Darrell, T.: Deep domain confusion: Maximizing for domain invariance. arXiv preprint arXiv:1412.3474 (2014) 10, 11

    Pith/arXiv arXiv 2014

  47. [47]

    IEEE transactions on knowledge and data engineering35(8), 8052–8072 (2022) 4

    Wang, J., Lan, C., Liu, C., Ouyang, Y., Qin, T., Lu, W., Chen, Y., Zeng, W., Yu, P.S.: Generalizing to unseen domains: A survey on domain generalization. IEEE transactions on knowledge and data engineering35(8), 8052–8072 (2022) 4

    2022

  48. [48]

    In: Proceedings of the AAAI conference on artificial intelligence

    Wang, J., Chen, B., Kang, B., Li, Y., Xian, W., Chen, Y., Xu, Y.: Ov-dquo: Open- vocabulary detr with denoising text query training and open-world unknown ob- jects supervision. In: Proceedings of the AAAI conference on artificial intelligence. vol. 39, pp. 7762–7770 (2025) 2, 4, 6, 10, 11

    2025

  49. [49]

    arXiv preprint arXiv:2502.00604 (2025) 14

    Wang, S., Bhartari, A.K., Li, B., Perdikaris, P.: Gradient alignment in physics- informed neural networks: A second-order optimization perspective. arXiv preprint arXiv:2502.00604 (2025) 14

    arXiv 2025

  50. [50]

    arXiv preprint arXiv:2108.10904 (2021) 1

    Wang, Z., Yu, J., Yu, A.W., Dai, Z., Tsvetkov, Y., Cao, Y.: Simvlm: Sim- ple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904 (2021) 1

    Pith/arXiv arXiv 2021

  51. [51]

    In: Proceedings of the 32nd ACM International Conference on Multimedia

    Wang,Z.,Zhou,W.,Xu,J.,Peng,Y.:Sia-ovd:Shape-invariantadapterforbridging the image-region gap in open-vocabulary detection. In: Proceedings of the 32nd ACM International Conference on Multimedia. pp. 4986–4994 (2024) 2, 4, 6, 10

    2024

  52. [52]

    In: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition

    Wu, S., Zhang, W., Jin, S., Liu, W., Loy, C.C.: Aligning bag of regions for open- vocabulary object detection. In: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition. pp. 15254–15264 (2023) 2, 3, 4, 5, 6, 9, 10

    2023

  53. [53]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Wu, X., Zhu, F., Zhao, R., Li, H.: Cora: Adapting clip for open-vocabulary de- tection with region prompting and anchor pre-matching. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7031–7040 (2023) 2, 4, 6, 10

    2023

  54. [54]

    In: European conference on computer vision

    Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in refer- ring expressions. In: European conference on computer vision. pp. 69–85. Springer (2016) 4

    2016

  55. [55]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14393–14402 (2021) 2, 3, 4, 9, 10, 11

    2021

  56. [56]

    arXiv preprint arXiv:1710.09412 (2017) 10

    Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017) 10

    Pith/arXiv arXiv 2017

  57. [57]

    In: European conference on computer vision

    Zhao, S., Zhang, Z., Schulter, S., Zhao, L., Vijay Kumar, B., Stathopoulos, A., Chandraker, M., Metaxas, D.N.: Exploiting unlabeled data with vision and lan- guage models for object detection. In: European conference on computer vision. pp. 159–175. Springer (2022) 4, 6, 10

    2022

  58. [58]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition

    Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition. pp. 16793–16803 (2022) 2, 4, 6, 11

    2022

  59. [59]

    In: European conference on com- puter vision

    Zhou, X., Girdhar, R., Joulin, A., Krähenbühl, P., Misra, I.: Detecting twenty- thousand classes using image-level supervision. In: European conference on com- puter vision. pp. 350–368. Springer (2022) 2, 3, 4, 5

    2022

  60. [60]

    International Journal of Computer Vision132(9), 3375–3407 (2024) 2

    Zhu, L., Yin, W., Yang, Y., Wu, F., Zeng, Z., Gu, Q., Wang, X., Zhou, C., Ye, N.: Vision-language alignment learning under affinity and divergence principles for few-shot out-of-distribution generalization. International Journal of Computer Vision132(9), 3375–3407 (2024) 2

    2024