Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

This paper claims that vision-language object detectors can be adapted at test time with a low-overhead combination of IoU-weighted entropy and image-conditioned prompt selection.

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

2026-08-04 13:23 UTC pith:RLLJCFG7

load-bearing objection A solid, honest first TTA method for VLODs, with broad benchmarks; the main caveats are missing variance and an unisolated IWE on driving scenes. the 3 major comments →

arxiv 2510.00458 v4 pith:RLLJCFG7 submitted 2025-10-01 cs.CV

VLOD-TTA: Test-Time Adaptation of Vision-Language Object Detectors

Atif Belal , Heitor R. Medeiros , Marco Pedersoli , Eric Granger This is my paper
classification cs.CV
keywords test-time adaptationvision-language object detectionIoU-weighted entropyprompt selectiondistribution shiftzero-shot detectionYOLO-WorldGrounding DINO
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.

This paper sets out to bring test-time adaptation to vision-language object detectors, a setting where the only prior method is a heavy mean-teacher approach. It claims that a detector's own dense, overlapping proposals can be used as free test-time signal: weighting an entropy loss by local IoU overlap steers adaptation toward spatially consistent clusters and away from isolated false-positive boxes. It adds image-conditioned prompt selection, which ranks candidate text prompts per image and fuses only the most compatible ones with the detector logits, instead of averaging all prompts. On YOLO-World and Grounding DINO, the combined method is reported to improve over zero-shot and all tested baselines across six datasets and fifteen corruptions, with one adaptation step per image and small parameter overhead. If right, this gives a practical, low-latency recipe for keeping open-vocabulary detectors accurate when the test domain drifts.

Core claim

The central discovery is that the spatial redundancy of a vision-language detector's own proposal output is a reliable correctness prior under domain shift, and that text prompts can be adapted per image by ranking them against region features. The paper repurposes the entropy-minimization objective used in classification TTA: instead of treating each proposal equally, it builds class-specific IoU graphs, weights each proposal by the size of its connected component raised to a power γ, and minimizes the weighted entropy. In parallel, it replaces prompt averaging with a selection of the top fraction ρ of prompts per class based on mean proposal-level similarity, then combines the selected-pro

What carries the argument

Two interlocking mechanisms carry the argument. The first, IoU-weighted entropy (IWE), constructs a per-class graph whose edges connect proposals with IoU above a threshold θ, defines connected components as proposal clusters, and assigns each proposal weight |component|^γ inside the entropy loss; this concentrates adaptation on dense, class-consistent regions. The second, image-conditioned prompt selection (IPS), scores each of T text prompts per class by its mean cosine similarity to normalized region features, retains the top-ρ prompts per class, averages only those, and fuses the result with the original detection score via λ. Lightweight adapters (vision and neck for YOLO-World, text en

Load-bearing premise

The method's core assumption is that regions with many mutually overlapping proposals that predict the same class are more likely to be correct, while isolated proposals are likely wrong; if a domain shift makes proposal overlap a poor proxy for correctness—such as scenes full of tiny, sparsely overlapping objects—the IoU-weighted objective loses its advantage.

What would settle it

Compare VLOD-TTA against standard entropy minimization (γ=0) on a target set of small-object urban scenes where proposals rarely overlap; if the IoU-weighted variant does not beat the unweighted baseline on such data, the overlap assumption is not holding. An even cleaner test: at test time, drop all but one proposal per object before adaptation and check whether the reported gains vanish.

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

If this is right

  • A single gradient step per image, with adapters reset afterward, is enough to improve both CNN- and transformer-based VLODs under a wide range of shifts.
  • IoU-weighted entropy reduces the confirmation bias of standard entropy minimization, cutting false positives from isolated boxes and improving AP50 and AP75, not just classification.
  • Per-image prompt selection is a lightweight alternative to prompt ensembling that avoids the score dilution prompt averaging can cause in detectors.
  • Gains hold across style, weather, low light, and 15 corruption types, including large relative improvements on digital corruptions such as JPEG compression and glass blur.
  • Because the method preserves zero-shot weights, it can be applied repeatedly without accumulating drift or forgetting source knowledge.

Where Pith is reading between the lines

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

  • The overlap-weighting principle is not specific to detection; the same IoU-graph weighting could be transferred to test-time adaptation for instance segmentation or other dense-prediction tasks that produce redundant output masks.
  • The paper resets adapters after every image; retaining a small per-domain memory of prompt rankings or cluster statistics could yield larger gains on long video streams, at the cost of latency.
  • The prompt-selection step effectively performs a per-image nearest-prototype lookup in embedding space; a natural extension is to cluster test images by prompt affinity and share selected prompts across a batch.
  • If overlap structure degrades under severe noise, a learned estimator of proposal reliability could replace the fixed IoU weight rather than resorting to standard entropy.

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

3 major / 4 minor

Summary. The paper proposes VLOD-TTA, a test-time adaptation method specifically for vision-language object detectors (YOLO-World and Grounding DINO). The method has two main components: (i) IoU-weighted entropy minimization (IWE), which weights each proposal's entropy by the size of its connected component in a class-conditional IoU graph, and (ii) image-conditioned prompt selection (IPS), which ranks a pool of prompts per class by mean image-level similarity and fuses the top-ranked prompts' logits with the detector's original logits. Only lightweight adapters and a text residual are optimized for a single step per image, then reset. Experiments cover stylized datasets (Watercolor, ClipArt, Comic), driving scenes (Cityscapes, BDD100K), low-light images (ExDark), and common corruptions (PASCAL-C, COCO-C). The paper reports consistent improvements over zero-shot and several TTA baselines adapted to detection.

Significance. If the empirical claims hold, this is a practical first TTA framework for VLODs: it adds only ~1.6M trainable parameters and runs at 20 FPS, notably cheaper than prompt-tuning or cache-based baselines. The benchmark is broad, spanning six main datasets plus 15 corruptions at multiple severities, and code is released. The proposed components are simple, well motivated, and architecture-agnostic. The main reservations are that the effectiveness of IWE is only isolated on stylized datasets, the style-shift results are obtained with hyperparameters tuned on the same test sets, and no statistical uncertainty is reported for the small gains on driving scenes. These issues weaken the 'consistently outperforms' claim as currently written.

major comments (3)
  1. [§4.2, Table 1; §5] The Conclusion states that 'IoU-weighted entropy can underperform in scenes dominated by many tiny, low-overlap objects (e.g., Cityscapes).' Yet the only ablation that isolates IWE from the rest of the pipeline (the γ sweep) is reported in Fig. 6 on Watercolor, ClipArt, and Comic. On Cityscapes and BDD100K, the full method always couples IWE with IPS and adapters, so the small margins over the Adapter baseline (e.g., YW Cityscapes mAP 19.4 vs 19.1, AP75 18.6 vs 18.3) cannot be attributed to IWE. The central mechanism's benefit in the exact regime the authors flag as problematic is therefore unverified. A γ=0 ablation (IWE off) on Cityscapes, BDD100K, and ExDark would resolve whether IWE helps, hurts, or is neutral in those domains, and is needed to support the cross-domain consistency claim.
  2. [§4.3–4.4, Fig. 6] Hyperparameters θ, γ, ρ, λ, and M are selected using sensitivity analyses on the same three style-shift datasets (Watercolor, ClipArt, Comic) that form the main style-shift results in Table 1. The reported style-shift gains are therefore in-sample with respect to hyperparameter selection. No sensitivity or validation is provided on driving, low-light, or corruption datasets. To make the robustness claim credible, either report sensitivity on held-out domains or justify fixed hyperparameters using a validation split.
  3. [§4.2, Tables 1–2] No error bars, standard deviations, or significance tests are reported. Several key improvements over the strongest baseline are small, e.g., YW Cityscapes AP75 18.6 vs 18.3, YW Cityscapes mAP 19.4 vs 19.1, and GD Watercolor mAP 38.9 vs 38.4. Because the abstract claims consistent outperformance across all shift types, the reader cannot judge whether these differences are within noise. Please report variability across seeds or bootstrap resamples, and where feasible paired significance tests.
minor comments (4)
  1. [Appendix A.1] The cosine–Euclidean equivalence is a trivial identity given ℓ2-normalized vectors and does not affect the method. It can be removed or compressed to a footnote.
  2. [§4.3 and Appendix A.5] The main text refers to 'Fig. 4/9 vary prompt strategies,' but Fig. 4 compares prompt-generation strategies while Fig. 9 (Appendix A.5) compares prompt averaging vs. selection. Clarify the references to avoid confusion.
  3. [Abstract] The phrase 'prior state-of-the-art method' is vague. The baselines in the paper are adaptations of classification TTA methods to detection, not prior VLOD-TTA methods; please specify what the prior SOTA refers to.
  4. [§4.1] The claim of '96 distinct test scenarios' is not broken down. A short enumeration (e.g., 6 datasets × 1 setting + 15 corruptions × 5 severities + ...) would make the count verifiable.

Circularity Check

0 steps flagged

No significant circularity: VLOD-TTA is an empirical TTA method validated on external benchmarks; its self-referential objectives are standard self-training, not hidden a priori derivations.

full rationale

The paper's contributions are empirical: two test-time objectives (IWE and IPS) plus adapters, benchmarked against external datasets (Watercolor/ClipArt/Comic, Cityscapes, BDD100K, ExDark, PASCAL-C, COCO-C) and TTA baselines. IWE weights proposal entropies by connected-component size in a class-consistent IoU graph (Eq. 3). This is defined in terms of the model's own posteriors, but the paper never presents it as an externally derived prediction; self-referential weighting is the standard mechanism of self-training TTA, and the claimed improvements are measured against held-out benchmarks, not derived from the objective's definition. IPS ranks prompts by mean class-specific logits and fuses selected scores (Eq. 5); again this is a bootstrap procedure whose value is shown empirically. The Appendix A.1 cosine-Euclidean proposition is a straightforward identity (||v-e||^2 = 2-2r) and does not constitute a derivation of the method. The only overlapping-author citation, Medeiros et al. (2025), is used for context ('the only work that adapts across modalities, but in a supervised setting') and to motivate visual prompting; it is not load-bearing for the central claim. The admitted weakness on tiny low-overlap objects (Cityscapes, Sec. 5) and the absence of a gamma=0 ablation there are concerns about attribution and experimental coverage, not circularity. Hyperparameters are tuned on style-shift datasets, which may bias those results, but this is not a definitional or self-citation circularity.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 0 invented entities

The method introduces no new physical or architectural entity beyond standard adapters and a residual prompt vector. Its load-bearing assumptions are domain-level: proposal overlap is a reliable spatial prior, entropy minimization is a safe self-training signal, and the chosen prompt pool is informative across all test shifts. The many hand-tuned hyperparameters (theta, gamma, rho, lambda, M, adapter sizes, prompt set) are the main unrecognized cost of the method.

free parameters (8)
  • IoU threshold theta = not stated in main text; sensitivity suggests 0.5-0.7
    Defines edges in the class-specific IoU graph (Sec 3.2). The value used in the main experiments is not reported in Sec 4.1.
  • Cluster exponent gamma = 1.1
    Controls component-size weighting w_i = |C(i)|^gamma; chosen via sensitivity runs on style-shift datasets (Fig 6b).
  • Prompt-selection fraction rho = 0.25
    Top fraction of prompts retained per class; tuned on style-shift datasets (Fig 6d).
  • Fusion coefficient lambda = 0.3 (YOLO-World), 0.1 (Grounding DINO)
    Blends selected-prompt scores with original detector scores in Eq. (5); chosen separately per detector from sensitivity analysis (Fig 6c).
  • Top-M proposals = 600
    Proposals retained to build the IoU graph; tuned (Fig 8).
  • Adapter sizes = reduction=4, kernel=3 (YW); reduction=16 (GD)
    Determines the number of trainable parameters; architecture choice not derived from the problem statement.
  • Prompt pool size and content = T=16 GPT-generated prompts per class
    Prompt content affects IPS and results; the ablation shows dataset-specific prompts give larger gains, so the dataset-agnostic GPT prompt set is a consequential design choice.
  • Test-time optimized adapters Phi and residual vector Delta = zero-initialized, optimized per image by IWE
    The adapter parameters and residual text vector are trained on each test image and reset afterward; they are the primary mechanism of adaptation.
axioms (5)
  • domain assumption Spatially overlapping, class-consistent proposals are more reliable than isolated boxes (Sec 3.2)
    Justifies weighting entropy by IoU-component size. If domain shift breaks proposal-overlap statistics, IWE loses its selectivity.
  • domain assumption Minimizing predictive entropy sharpens correct labels under shift and does not systematically amplify false positives once weighted (Sec 3.1, Sec 5)
    Standard TTA premise inherited from TENT/TPT; the authors acknowledge it underperforms on tiny low-overlap objects.
  • domain assumption Lightweight adapters with frozen backbone and per-image reset preserve zero-shot capability (Sec 3.4)
    Assumes no catastrophic drift in a single adaptation step; supported only by empirical results, not proven.
  • domain assumption Mean proposal-level cosine similarity is a valid image-level prompt-compatibility score (Sec 3.3, A.1)
    IPS ranks prompts by average similarity; the Cosine-Euclidean proposition is an identity, but the ranking's validity is assumed.
  • ad hoc to paper GPT-generated, dataset-agnostic prompts are an unbiased prompt pool for all test domains (Sec A.6)
    Prompt content is a design choice; the ablation shows dataset-specific prompts give larger gains, so the prompt-generation procedure is consequential.

pith-pipeline@v1.3.0-alltime-deepseek · 25452 in / 11211 out tokens · 273713 ms · 2026-08-04T13:23:12.146384+00:00 · methodology

0 comments
read the original abstract

Vision-language object detectors (VLODs) such as YOLO-World and Grounding DINO exhibit strong zero-shot generalization, but their performance degrades under distribution shift. Test-time adaptation (TTA) offers a practical way to adapt models during inference using only unlabeled target (test) data. However, while TTA has made substantial progress in vision-language classification, its application to VLODs remains largely unexplored. The only prior method relies on a mean-teacher framework that introduces significant latency and memory overhead. To this end, we introduce VLOD-TTA, a TTA method that leverages dense proposal overlap and image-conditioned prompts to adapt VLODs with low additional overhead. VLOD-TTA combines (i) an IoU-weighted entropy objective that emphasizes spatially coherent proposal clusters and mitigates confirmation bias from isolated boxes, and (ii) image-conditioned prompt selection that ranks prompts by image-level compatibility and aggregates the most informative prompt scores for detection. Our experiments across diverse distribution shifts, including artistic domains, adverse driving conditions, low-light imagery, and common corruptions, indicate that VLOD-TTA consistently outperforms standard TTA baselines and the prior state-of-the-art method using YOLO-World and Grounding DINO. Our code: https://github.com/imatif17/VLOD-TTA

Figures

Figures reproduced from arXiv: 2510.00458 by Atif Belal, Eric Granger, Heitor R. Medeiros, Marco Pedersoli.

Figure 1
Figure 1. Figure 1: Motivation (IWE). Left→right: (i) proposals from two classes—Person (red) (cluster size = 167, max score = 0.14) and Dog (blue) (cluster size = 45, max score = 0.15); (ii) ZS scores fall below the threshold, resulting in a missed detection; (iii) standard entropy minimization over￾confidently produces a dog false positive; and (iv) our IoU-weighted entropy minimization focuses updates on dense clusters and… view at source ↗
Figure 2
Figure 2. Figure 2: Motivation (IPS). Left→right: (i) ZS predictions with a correct detection Person (red) and a false positive Dog (blue); (ii) prompt–class score heatmap with circles marking prompts se￾lected by our image-conditioned strategy and right-margin bars showing SPS − SPA; (iii) prompt averaging (PA) reduces the class score, producing no detections; and (iv) prompt selection (PS) sup￾presses the dog false positive… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of our VLOD-TTA. Given an input image and a set of prompts, the text encoder produces embeddings that interact with region proposals via the vision–language head to compute similarity scores. IPS performs top-ρ prompt selection and averages the selected prompts to obtain per-proposal class scores. Then, it combines per-proposal entropy with IoU-based weights to form an IWE objective that drives ro… view at source ↗
Figure 6
Figure 6. Figure 6: Variation in performance with hyperparameters on three style-shift datasets. The IoU threshold (θ) and IoU-cluster exponent (γ) influence the IWE, while the fusion coefficient (λ) and selection fraction (ρ) are IPS hyperparameters. 4.3 ABLATION STUDIES [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 4
Figure 4. Figure 4: Prompt generation strategies. ∆mAP50 over three style-shift datasets mea￾sured relative to ZS. Variation in performance with different prompt-generation strategies. GPT-generated prompts were used without any dataset information. In this ablation, we compare two alternatives, namely dataset-specific GPT prompts and CLIP-style prompts (see Sec. A.6). The improvement over ZS on Watercolor, ClipArt, and Comic… view at source ↗
Figure 5
Figure 5. Figure 5: Adapters in differ￾ent detector modules. Mean ∆mAP50 averaged over three style-shift datasets, relative to ZS. Effect of adapter placement across modules. In [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: YOLO-World detections across different approaches: Each column corresponds to a different method: (a) GT (Ground Truth), (b) ZS (Zero-Shot), (c) Adapter, and (d) VLOD-TTA. Each color represents a different object category. Effect of γ for graph construction. The exponent γ controls how strongly component size influ￾ences the IoU-weighted entropy. When γ=0, all wi are equal and the objective reduces to stan… view at source ↗
Figure 8
Figure 8. Figure 8: Effect of top-M proposals on three style-shift datasets. IWE uses the top-M proposals to construct the IoU graph [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison of Prompt Averaging and Prompt Selection. ∆mAP50 over three style-shift datasets measured relative to ZS. Method FPS Train Params (M) ZS 89 0.00 FFT 89 76.81 TPT 9 1.12 VPT 18 3.93 DPE 15 0.31 Adapter 22 1.52 Ours 20 1.61 [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Effect of batch size and number of augmentations on performance. We report mAP50 across three style-shift datasets on YOLO-World. YOLO-World-L Watercolor ClipArt Comic Method mAP AP50 AP75 mAP AP50 AP75 mAP AP50 AP75 ZS 32.8 55.3 33.0 31.1 50.6 32.6 23.3 37.9 23.6 Ours 34.1 58.3 34.2 33.2 53.9 34.3 26.5 42.8 27.2 Grounding DINO-B Watercolor ClipArt Comic Method mAP AP50 AP75 mAP AP50 AP75 mAP AP50 AP75 ZS… view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative comparison of standard and IoU-weighted entropy. Top row: ground truth and predictions using standard and IoU-weighted entropy. Bottom row: heatmaps of the IoU graph clusters from the ZS model. For each cluster, the predicted category, cluster size, and maxi￾mum score are displayed. Only the top four clusters are shown. standard entropy tends to raise bird scores while the person class’s score… view at source ↗
Figure 12
Figure 12. Figure 12: YOLO-World detections across different approaches: Each column corresponds to a different approach: (a) GT (Ground Truth), (b) ZS (Zero-Shot), (c) Adapter, and (d) VLOD-TTA. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Reward-Guided Semantic Evolution for Test-time Adaptive Object Detection

    cs.CV 2026-05 unverdicted novelty 6.0

    RGSE adapts text embeddings at test time via evolutionary search, using cosine similarity rewards from high-confidence visual proposals to improve open-vocabulary object detection under distribution shifts.

Reference graph

Works this paper leans on

42 extracted references · 23 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Exploring test-time adaptation for object detection in continually changing environments, 2025

    Shilei Cao, Juepeng Zheng, Yan Liu, Baoquan Zhao, Ziqi Yuan, Weijia Li, Runmin Dong, and Haohuan Fu. Exploring test-time adaptation for object detection in continually changing environments, 2025. URL https://arxiv.org/abs/2406.16439

    Pith/arXiv arXiv 2025

  2. [2]

    Conv-adapter: Exploring parameter efficient transfer learning for convnets, 2024

    Hao Chen, Ran Tao, Han Zhang, Yidong Wang, Xiang Li, Wei Ye, Jindong Wang, Guosheng Hu, and Marios Savvides. Conv-adapter: Exploring parameter efficient transfer learning for convnets, 2024. URL https://arxiv.org/abs/2208.07463

    Pith/arXiv arXiv 2024

  3. [3]

    Stfar: Improving object detection robustness at test-time by self-training with feature alignment regularization, 2023

    Yijin Chen, Xun Xu, Yongyi Su, and Kui Jia. Stfar: Improving object detection robustness at test-time by self-training with feature alignment regularization, 2023. URL https://arxiv.org/abs/2303.17937

    Pith/arXiv arXiv 2023

  4. [4]

    Yolo-world: Real-time open-vocabulary object detection, 2024

    Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection, 2024. URL https://arxiv.org/abs/2401.17270

    Pith/arXiv arXiv 2024

  5. [5]

    The cityscapes dataset for semantic urban scene understanding, 2016

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding, 2016. URL https://arxiv.org/abs/1604.01685

    Pith/arXiv arXiv 2016

  6. [6]

    Williams, John Winn, and Andrew Zisserman

    Mark Everingham, Luc Gool, Christopher K. Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. Int. J. Comput. Vision, 88 0 (2): 0 303–338, June 2010. ISSN 0920-5691. doi:10.1007/s11263-009-0275-4. URL https://doi.org/10.1007/s11263-009-0275-4

    doi:10.1007/s11263-009-0275-4 2010

  7. [7]

    Frustratingly easy test-time adaptation of vision-language models

    Matteo Farina, Gianni Franchi, Giovanni Iacca, Massimiliano Mancini, and Elisa Ricci. Frustratingly easy test-time adaptation of vision-language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=eQ6VjBhevn

    2024

  8. [8]

    Diverse data augmentation with diffusions for effective test-time prompt tuning, 2023

    Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning, 2023. URL https://arxiv.org/abs/2308.06038

    Pith/arXiv arXiv 2023

  9. [9]

    Deep learning-based object detection in augmented reality: A systematic review

    Yalda Ghasemi, Heejin Jeong, Sung Ho Choi, Kyeong-Beom Park, and Jae Yeol Lee. Deep learning-based object detection in augmented reality: A systematic review. Computers in Industry, 139: 0 103661, 2022

    2022

  10. [10]

    Open-vocabulary object detection via vision and language knowledge distillation, 2022

    Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation, 2022. URL https://arxiv.org/abs/2104.13921

    Pith/arXiv arXiv 2022

  11. [11]

    Deep learning for object detection and scene perception in self-driving cars: Survey, challenges, and open issues

    Abhishek Gupta, Alagan Anpalagan, Ling Guan, and Ahmed Shaharyar Khwaja. Deep learning for object detection and scene perception in self-driving cars: Survey, challenges, and open issues. Array, 10: 0 100057, 2021

    2021

  12. [12]

    Benchmarking neural network robustness to common corruptions and perturbations

    Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019

    2019

  13. [13]

    Parameter-efficient transfer learning for nlp, 2019

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp, 2019. URL https://arxiv.org/abs/1902.00751

    Pith/arXiv arXiv 2019

  14. [14]

    Cross-domain weakly-supervised object detection through progressive domain adaptation, 2018

    Naoto Inoue, Ryosuke Furuta, Toshihiko Yamasaki, and Kiyoharu Aizawa. Cross-domain weakly-supervised object detection through progressive domain adaptation, 2018. URL https://arxiv.org/abs/1803.11365

    Pith/arXiv arXiv 2018

  15. [15]

    Visual prompt tuning

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision (ECCV), 2022

    2022

  16. [16]

    Efficient test-time adaptation of vision-language models

    Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. The IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

    2024

  17. [17]

    Grounded language-image pre-training, 2022

    Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training, 2022. URL https://arxiv.org/abs/2112.03857

    Pith/arXiv arXiv 2022

  18. [18]

    Clu-cnns: Object detection for medical images

    Zhuoling Li, Minghui Dong, Shiping Wen, Xiang Hu, Pan Zhou, and Zhigang Zeng. Clu-cnns: Object detection for medical images. Neurocomputing, 350: 0 53--59, 2019

    2019

  19. [19]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll \'a r, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pp.\ 740--755. Springer, 2014

    2014

  20. [20]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection, 2024

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding dino: Marrying dino with grounded pre-training for open-set object detection, 2024. URL https://arxiv.org/abs/2303.05499

    Pith/arXiv arXiv 2024

  21. [21]

    Getting to know low-light images with the exclusively dark dataset

    Yuen Peng Loh and Chee Seng Chan. Getting to know low-light images with the exclusively dark dataset. Computer Vision and Image Understanding, 178: 0 30--42, 2019. doi:https://doi.org/10.1016/j.cviu.2018.10.010

    doi:10.1016/j.cviu.2018.10.010 2019

  22. [22]

    Medeiros, Atif Belal, Srikanth Muralidharan, Eric Granger, and Marco Pedersoli

    Heitor R. Medeiros, Atif Belal, Srikanth Muralidharan, Eric Granger, and Marco Pedersoli. Visual modality prompt for adapting vision-language object detectors, 2025. URL https://arxiv.org/abs/2412.00622

    Pith/arXiv arXiv 2025

  23. [23]

    Benchmarking robustness in object detection: Autonomous driving when winter is coming

    Claudio Michaelis, Benjamin Mitzkus, Robert Geirhos, Evgenia Rusak, Oliver Bringmann, Alexander S Ecker, Matthias Bethge, and Wieland Brendel. Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484, 2019

    Pith/arXiv arXiv 1907

  24. [24]

    Pawan Kumar Mishra and G. P. Saroha. A study on video surveillance system for object detection and tracking. In 2016 3rd International Conference on Computing for Sustainable Global Development (INDIACom), pp.\ 221--226, 2016

    2016

  25. [25]

    Learning transferable visual models from natural language supervision, 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. URL https://arxiv.org/abs/2103.00020

    Pith/arXiv arXiv 2021

  26. [26]

    Fully test-time adaptation for object detection

    Xiaoqian Ruan and Wei Tang. Fully test-time adaptation for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pp.\ 1038--1047, June 2024

    2024

  27. [27]

    Claude E. Shannon. A mathematical theory of communication. Bell System Technical Journal, 27 0 (3): 0 379--423, 1948

    1948

  28. [28]

    Objects365: A large-scale, high-quality dataset for object detection

    Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pp.\ 8429--8438, 2019. doi:10.1109/ICCV.2019.00852

    arXiv 2019

  29. [29]

    Test-time prompt tuning for zero-shot generalization in vision-language models, 2022

    Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models, 2022. URL https://arxiv.org/abs/2209.07511

    Pith/arXiv arXiv 2022

  30. [30]

    Tent: Fully test-time adaptation by entropy minimization, 2021

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization, 2021. URL https://arxiv.org/abs/2006.10726

    Pith/arXiv arXiv 2021

  31. [31]

    What, how, and when should object detectors update in continually changing test domains?, 2023

    Jayeon Yoo, Dongkwan Lee, Inseop Chung, Donghyun Kim, and Nojun Kwak. What, how, and when should object detectors update in continually changing test domains?, 2023. URL https://arxiv.org/abs/2312.08875

    Pith/arXiv arXiv 2023

  32. [32]

    Bdd100k: A diverse driving dataset for heterogeneous multitask learning, 2020

    Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning, 2020. URL https://arxiv.org/abs/1805.04687

    Pith/arXiv arXiv 2020

  33. [33]

    Open-vocabulary object detection using captions, 2021

    Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions, 2021. URL https://arxiv.org/abs/2011.10678

    Pith/arXiv arXiv 2021

  34. [34]

    Dual prototype evolving for test-time generalization of vision-language models, 2024

    Ce Zhang, Simon Stepputtis, Katia Sycara, and Yaqi Xie. Dual prototype evolving for test-time generalization of vision-language models, 2024. URL https://arxiv.org/abs/2410.12790

    Pith/arXiv arXiv 2024

  35. [35]

    Memo: Test time robustness via adaptation and augmentation, 2022

    Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation, 2022. URL https://arxiv.org/abs/2110.09506

    Pith/arXiv arXiv 2022

  36. [36]

    Regionclip: Region-based language-image pretraining, 2021

    Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. Regionclip: Region-based language-image pretraining, 2021. URL https://arxiv.org/abs/2112.09106

    Pith/arXiv arXiv 2021

  37. [37]

    Detecting twenty-thousand classes using image-level supervision, 2022

    Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krähenbühl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision, 2022. URL https://arxiv.org/abs/2201.02605

    Pith/arXiv arXiv 2022

  38. [38]

    Object detection in 20 years: A survey

    Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 111 0 (3): 0 257--276, 2023

    2023

  39. [39]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  40. [40]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

  41. [41]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

  42. [42]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...