Pith. sign in

REVIEW 3 major objections 5 minor 58 references

A test-time adaptation method for monocular 3D object detection that jointly minimizes semantic and geometric uncertainty, using a label-free conjugate focal loss and a semantic-guided normal consistency loss, recovers detection accuracy un

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 ↗

DUO adapts monocular 3D detectors at test time by combining a label-free conjugate focal loss for semantic uncertainty with a semantic-guided normal-field consistency loss for geometric uncertainty.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Consistent empirical gains for a new TTA loss in monocular 3D detection, but the core derivation is not rigorous; treat the method as a heuristic. the 3 major comments →

arxiv 2508.20488 v1 pith:KIKCJGTO submitted 2025-08-28 cs.CV

Adaptive Dual Uncertainty Optimization: Boosting Monocular 3D Object Detection under Test-Time Shifts

Zixuan Hu , Dongxiao Li , Xinzhu Ma , Shixiang Tang , Xiaotong Li , Wenhan Yang , Ling-Yu Duan This is my paper
classification cs.CV
keywords test-time adaptationmonocular 3D object detectionuncertainty minimizationconjugate focal lossnormal consistency lossdomain shiftKITTI-CnuScenes
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.

The reading

DUO is a fully test-time adaptation (TTA) method for monocular 3D object detection that targets the two uncertainties that grow together under domain shifts: semantic uncertainty (ambiguous class predictions) and geometric uncertainty (unstable spatial and depth localization). The paper's central claim is that jointly minimizing both, rather than entropy alone, is what lets a pretrained detector adapt online to corrupted or weather-shifted inputs. To achieve this, it derives a Conjugate Focal Loss from a Legendre–Fenchel reformulation of the focal loss, which reweights predictions without ground-truth labels and draws training toward low-score objects that entropy minimization neglects. Alongside it, a semantic-guided normal consistency loss smooths the predicted depth map's surface normals in confident regions, stabilizing 3D geometry and feeding back into better classification. The paper reports consistent gains on KITTI-C's 13 corruptions and on nuScenes day/night and sunny/rainy tasks across two different base detectors.

Core claim

From the focal loss's convex-conjugate structure, the paper derives a new label-free loss, the Conjugate Focal Loss (CFL): L_CFL(x) = -α(1-p)^γ (I + γ(1 - log p) p p^T - γ log p diag(p))^{-1} p log p. The derived pseudo-label y0 = (I + γ(1-log p)p p^T - γ log p diag(p))^{-1} p dynamically weights every class using only the prediction distribution, so low-confidence objects receive larger adaptation gradients without requiring ground-truth labels. On the geometric side, DUO converts the predicted depth map into a normal field via Sobel operators and penalizes local normal inconsistency with an edge-aware loss, masked to regions where the semantic branch is confident. The paper claims these tw

What carries the argument

The load-bearing object is the Legendre–Fenchel conjugate of the focal loss. The paper rewrites the focal loss as L_FL = f(h) - y^T g(h) with f(h) = α log s and g(h) = αh + α((1-p)^γ - 1) log p, then invokes the convex-duality condition that a conjugate exists when g is locally invertible. A second-order Taylor approximation yields the explicit pseudo-label y0 ≈ (I + γ(1-log p)p p^T - γ log p diag(p))^{-1} p, which defines the Conjugate Focal Loss for unlabeled data. The other machinery is the normal field constraint: depth map to Sobel gradients to surface normals, an edge-aware smoothness penalty, and a semantic-confidence mask that limits the geometric regularizer to reliable regions.

Load-bearing premise

The entire self-training loss rests on the assumption that the mapping g(h) between the model's logits and the dual coordinates stays locally invertible on the shifted test predictions, and that the second-order Taylor approximation used to write down the conjugate pseudo-label is accurate for every detected object; if either fails on a meaningful fraction of objects, the 'conjugate' loss degenerates into an arbitrary entropy reweighting.

What would settle it

Compute the Jacobian ∇h g(h) for every detected object on a shifted validation set (e.g., KITTI-C Gaussian noise at severity 5) using the pretrained detector's logits. If a substantial fraction of objects have ∇h g nearly singular (smallest singular value close to zero), and if masking those objects out changes the reported AP advantages, then the CFL derivation is not what is driving the improvement. Alternatively, replace CFL with a random per-class entropy reweighting matched in magnitude and test whether the +2.2 AP advantage disappears.

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

If this is right

  • Monocular 3D detectors can be adapted online to corrupted inputs with no labels and no source data, improving AP on all 13 KITTI-C corruption types tested.
  • Low-score and high-uncertainty objects receive effective training signal, countering the foreground-background imbalance that plain entropy minimization worsens in detection.
  • The multi-head depth estimator avoids collapsing into a single deterministic head, because the normal-field constraint keeps geometric uncertainty balanced across heads.
  • The framework transfers across two distinct base detectors (MonoFlex and MonoGround) and across simulated corruptions as well as real day/night and sunny/rainy shifts.
  • The hyperparameters α and γ carry over from source training, so adaptation requires no extra hyperparameter tuning beyond the default settings.

Where Pith is reading between the lines

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

  • The conjugate-focal-loss trick is not specific to 3D detection: any dense detector or segmentation model trained with focal loss could adopt the same label-agnostic reweighting when adapting to unlabeled target data.
  • The normal-consistency term is a generic geometric regularizer and could be combined with other monocular depth or bird's-eye-view perception losses, possibly improving robustness beyond the two tested detectors.
  • The paper's own invertibility argument gives a cheap diagnostic: if a shifted prediction makes the matrix I + γ(1-log p)p p^T - γ log p diag(p) near-singular, CFL's behavior degenerates; masking or down-weighting such objects could be a testable safeguard.
  • The complementary semantic-geometric loop suggests a natural extension to video or multi-frame TTA, where temporal geometry could further stabilize the depth normal field while semantic confidence guides which frames to trust.
Share X Bluesky LinkedIn Reddit HN

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 / 5 minor

Summary. The paper proposes DUO, a test-time adaptation method for monocular 3D object detection that jointly minimizes semantic and geometric uncertainties. The semantic branch introduces a Conjugate Focal Loss (CFL), derived by casting the focal loss in a Legendre-Fenchel form and using a local invertibility and Taylor approximation to obtain a label-free pseudo-label, while the geometric branch uses a Sobel-based normal consistency loss guided by low-semantic-uncertainty regions. The method is evaluated on KITTI-C (13 corruptions, multiple severities) and nuScenes day/night and sunny/rainy tasks, with two base detectors, and it reports consistent improvements over TENT, EATA, DeYO, and MonoTTA.

Significance. If the theoretical derivation were fully sound, DUO would be a noteworthy contribution: it is, to the best of the authors' claim, the first TTA framework for M3OD that explicitly targets both semantic and geometric uncertainty, and the empirical gains are consistent across two base models, two datasets, and multiple corruption types. The paper also provides code, severity-level ablations, component ablations, and runtime comparisons, which are valuable for reproducibility. However, the central novelty rests on the claim that CFL is a principled conjugate loss rather than an ad hoc reweighting of entropy. That claim is currently not secured by the supplied proof, so the practical contribution is strong but the theoretical framing needs substantial repair or honest re-characterization.

major comments (3)
  1. [Appendix A.2, Eq. (24)] The invertibility proof is incomplete. The text aims to show that g(h) = αh + α((1−p)^γ −1)log p is locally invertible, but the argument only establishes λ(∇_h g) ≥ 0, i.e., positive semidefiniteness. The inverse function theorem requires a non-singular Jacobian. The sentence 'H is full rank on its subspace, ensuring that ∇_h g is non-singular' does not follow from PSD-ness, since a PSD matrix can be singular on the orthogonal complement. Moreover, Eq. (23) bounds eigenvalues of the non-symmetric product D·H using products of λmin(D) and λmin(H); this is not generally valid for row-scaled non-symmetric matrices. The paper needs either a strict eigenvalue lower bound λ_min(∇_h g)>0 with explicit conditions on p, or a clear statement that invertibility is an empirical assumption backed by measurements on test predictions.
  2. [Sec. 5.1 / Appendix A.3, Eqs. (34)-(35)] The step from Eq. (33) to Eq. (35) replaces the exact Jacobian by a truncated Taylor expansion, discarding higher-order terms in p and log p without any error bound or neighborhood estimate. The resulting matrix M = I + γ(1−log p) pp^T − γ log p diag(p) is then asserted to be the approximate Jacobian. If M is singular or ill-conditioned for a non-negligible fraction of test predictions, y0 = M^{-1}p is not a valid conjugate pseudo-label and L_CFL degenerates to an ad hoc reweighting of entropy. This is load-bearing because the 'derived' status of CFL is the paper's main claimed novelty. Please either provide a quantitative error bound, report the eigenvalue/condition-number statistics of M over the test set, or explicitly present CFL as a heuristic approximation rather than as a conjugate loss.
  3. [Sec. 5.1, Eq. (7)] The central loss equation is not well-formed as written. The left-hand side L_CFL(x) is a scalar, while the right-hand side contains the vector p log p (elementwise log) multiplied by the matrix M^{-1}, which yields a vector. No scalar reduction (e.g., a dot product with y0, a sum, or a norm) is specified. Additionally, the term p^T p in Eq. (7) should presumably be the outer product pp^T as in Eq. (6), otherwise the matrix dimensions are incompatible. Because this equation defines the training loss, the implementation cannot be reproduced from the text as it stands. Please correct the notation and specify the exact scalar form used in the experiments.
minor comments (5)
  1. [Sec. 5.1, Eq. (6)] The notation y0 ≜ ∇_h(f∘g^{-1}) / ∇_h z |_{z=g(h)} is confusing. The numerator and denominator are not a clean quotient; this should be written as y0 = (∇_h g(h0))^{-1} ∇_h f(h0) at the relevant point.
  2. [Appendix A.2, Eq. (17)] The Jacobian expression divides by p_i. For degenerate predictions where some p_i approaches 0, the expression is unbounded. The paper should state the assumed range of p (e.g., regularized softmax probabilities) or handle this case explicitly.
  3. [Related Work / Footnote 1] The derivation in Sec. 5.1 closely follows TAPL [12], and the paper cites it, but the text should explicitly state which component is newly derived beyond applying the conjugate pseudo-label framework to focal loss. This matters for the novelty claim.
  4. [References] There are several typos in the reference list, e.g., 'V ora' in [5] and 'Internetional Conference' in [37]. Please proofread.
  5. [Tables 5-6] In Table 5, the Ours row contains a stray 'X' where the Gaussian value should be. This appears to be a LaTeX artifact and should be fixed before final submission.

Circularity Check

0 steps flagged

No significant circularity: the CFL derivation is a self-contained extension of the cited conjugate-pseudo-label framework, and the model-dependent pseudo-label is standard TTA self-training, not a circular reduction.

full rationale

The paper's central derivation (Sec. 5.1, App. A) starts from the algebraic focal-loss identity L_FL = f(h) - y^T g(h) (Eq. 3, App. Eq. 16), applies the Legendre-Fenchel conjugation, and estimates a label-agnostic pseudo-label y0 = (∇g)^{-1}∇f from the model's own softmax output p (Eq. 6, App. Eqs. 33-36). This is exactly the conjugate-pseudo-label construction introduced in the cited TAPL work [12]; using the model's own output as a pseudo-label is the standard TTA self-training loop, not a logical reduction of the result to its inputs. The loss L_CFL (Eq. 7) is an unsupervised objective defined over the model output, not a fitted parameter later reported as a prediction. The paper's self-citations [16,17] appear only in related work and do not carry the derivation or the novelty claims. The identified gaps are rigor/correctness issues rather than circularity: App. A.2 proves only positive semidefiniteness (Eq. 24) and then asserts non-singularity from non-degenerate p without a strict lower eigenvalue bound, and Eq. (34) drops higher-order Taylor terms without error bounds. These omissions undermine the proof that Eq. (7) is exactly the conjugate loss, but they do not make the derivation equivalent to its inputs. The empirical comparisons (Tables 1-3) are against external benchmarks with fixed base models, so the reported gains are self-contained evidence independent of the derivation. Overall, no circular step is present; the score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The method introduces no new physical entities. The free parameters are typical TTA hyperparameters; λ and β are tuned, while α and γ are inherited from source training. The strongest axiom is local invertibility of g, which is asserted rather than fully proven.

free parameters (4)
  • lambda (NCL weight) = 0.7
    Coefficient balancing conjugate focal loss and normal consistency loss in Eq. 14; tuned via ablation on Gaussian shift (Fig. 7a).
  • alpha (focal weight) = 4
    Weighting scale in CFL, set to source-phase focal loss default; Fig. 7b shows performance is robust but α=4 is chosen.
  • gamma (focal exponent) = 2
    Exponent in CFL, inherited from focal loss source training and used in the Legendre-Fenchel proof (Appendix A.2 assumes γ=2).
  • beta (moving average) = 0.1
    Moving-average factor for the semantic uncertainty threshold in Eq. 12; set by default, no sensitivity analysis.
axioms (6)
  • standard math Legendre-Fenchel conjugation applies to the focal loss reformulated as f(h)-y^T g(h)
    Used in Sec. 5.1 and Appendix A.1; convex analysis is standard but the reformulation itself is derived in the paper.
  • domain assumption The function g(h)=αh+α((1-p)^γ-1)log p is locally invertible at test-time predictions
    Required to define the conjugate pseudo-label y0; Appendix A.2 attempts a proof that only establishes positive semidefinite Jacobian.
  • domain assumption The source representation h is close to a local optimum h0
    Invoked in Sec. 5.1 Eq. 5 to replace the conjugate condition by gradient relations; cited to pre-training literature [9,35].
  • ad hoc to paper Higher-order terms in p and log p can be neglected in the Taylor expansion
    Appendix A.3 Eq. (34-35) drops terms without error bounds; this is the main approximation that turns the conjugate loss into Eq. 7.
  • domain assumption The depth map D produced by the multi-head depth branch is dense enough for Sobel-gradient normal estimation
    Sec. 5.2 computes normals from D restored to input resolution; if D is sparse or misaligned, the normal consistency loss is unreliable.
  • domain assumption Regions with low semantic uncertainty (by moving-average threshold) contain trustworthy geometry for the normal constraint
    Eq. 12-13 select boxes by semantic uncertainty to build the mask M; this assumes semantic confidence correlates with geometric reliability.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive Dual Uncertainty Optimization: Boosting Monocular 3D Object Detection under Test-Time Shifts." pith.science (2026). https://pith.science/paper/KIKCJGTO

@misc{pith2026250820488,
  author       = {Pith},
  title        = {Pith review of: Adaptive Dual Uncertainty Optimization: Boosting Monocular 3D Object Detection under Test-Time Shifts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KIKCJGTO}},
  note         = {Machine review of arXiv:2508.20488}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Accurate monocular 3D object detection (M3OD) is pivotal for safety-critical applications like autonomous driving, yet its reliability deteriorates significantly under real-world domain shifts caused by environmental or sensor variations. To address these shifts, Test-Time Adaptation (TTA) methods have emerged, enabling models to adapt to target distributions during inference. While prior TTA approaches recognize the positive correlation between low uncertainty and high generalization ability, they fail to address the dual uncertainty inherent to M3OD: semantic uncertainty (ambiguous class predictions) and geometric uncertainty (unstable spatial localization). To bridge this gap, we propose Dual Uncertainty Optimization (DUO), the first TTA framework designed to jointly minimize both uncertainties for robust M3OD. Through a convex optimization lens, we introduce an innovative convex structure of the focal loss and further derive a novel unsupervised version, enabling label-agnostic uncertainty weighting and balanced learning for high-uncertainty objects. In parallel, we design a semantic-aware normal field constraint that preserves geometric coherence in regions with clear semantic cues, reducing uncertainty from the unstable 3D representation. This dual-branch mechanism forms a complementary loop: enhanced spatial perception improves semantic classification, and robust semantic predictions further refine spatial understanding. Extensive experiments demonstrate the superiority of DUO over existing methods across various datasets and domain shift types.

Figures

Figures reproduced from arXiv: 2508.20488 by Dongxiao Li, Ling-Yu Duan, Shixiang Tang, Wenhan Yang, Xiaotong Li, Xinzhu Ma, Zixuan Hu.

Figure 1
Figure 1. Figure 1: (a) Illustration of M3OD. (b, c) Under test-time shifts, increased semantic and geometric uncertainties lead to a degrada￾tion of semantic perception and spatial localization capabilities. derstand 3D scenes directly from 2D images, as shown in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Semantic uncertainty (prediction entropy) and geometric uncertainty (average depth uncertainty) of the M3OD model under [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of our DUO method. (a) Meta-framework of the base models, where a backbone network connects to multi-branch predictors for estimating various object properties. (b) Through our theoretical conjugate optimization framework, we derive a conjugate focal loss LCFL that can adaptively adjust training weights across all classes without the ground-truth labels, while also identifying low￾semantic-uncerta… view at source ↗
Figure 4
Figure 4. Figure 4: Effects of components in our method. (a) shows the distribution of detection scores during adaptation with conjugate focal loss. (b) records the variation of depth uncertainty in differ￾ent heads during adaptation with normal field constraints. maintaining the best performance across all four scenarios, further demonstrating its effectiveness and superiority. 6.3. Ablation Study In this section, without lo… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative examples on KITTI-C. In each row, we pro [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: (a) Performance with varying strengths λ of the normal field constraint. (b) Performance with different weighting scare α of the conjugate focal loss [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: An illustration of 13 distinct types of corruptions in the severity level 3 of the KITTI-C dataset. [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: An illustration of the Daytime, Night, Sunny, and Rainy scenarios of the nuScenes dataset. [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] 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

58 extracted references · 54 canonical work pages

  1. [1]

    A survey on 3d object detection methods for autonomous driving applica- tions

    Eduardo Arnold, Omar Y Al-Jarrah, Mehrdad Dianati, Saber Fallah, David Oxtoby, and Alex Mouzakitis. A survey on 3d object detection methods for autonomous driving applica- tions. IEEE Transactions on Intelligent Transportation Sys- tems, 2019. 1

    work page 2019

  2. [2]

    A theory of learning from different domains.Machine learn- ing, 79:151–175, 2010

    Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains.Machine learn- ing, 79:151–175, 2010. 1

    work page 2010

  3. [3]

    Convex analysis and optimization

    Dimitri Bertsekas, Angelia Nedic, and Asuman Ozdaglar. Convex analysis and optimization. Athena Scientific, 2003. 2, 4

    work page 2003

  4. [4]

    Parameter-free online test-time adaptation

    Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 8344–8353, 2022. 2

    2022

  5. [5]

    nuscenes: A multi- modal dataset for autonomous driving

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom. nuscenes: A multi- modal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621–11631, 2020. 2, 6, 18

    work page 2020

  6. [6]

    Foreground- background imbalance problem in deep object detectors: A review

    Joya Chen, Qi Wu, Dong Liu, and Tong Xu. Foreground- background imbalance problem in deep object detectors: A review. In 2020 IEEE Conference on Multimedia Infor- mation Processing and Retrieval (MIPR) , pages 285–290. IEEE, 2020. 3

    work page 2020

  7. [7]

    Monopair: Monocular 3d object detection using pairwise spatial relationships

    Yongjian Chen, Lei Tai, Kai Sun, and Mingyang Li. Monopair: Monocular 3d object detection using pairwise spatial relationships. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 12093–12102, 2020. 1

    work page 2020

  8. [8]

    Learning depth-guided con- volutions for monocular 3d object detection

    Mingyu Ding, Yuqi Huo, Hongwei Yi, Zhe Wang, Jianping Shi, Zhiwu Lu, and Ping Luo. Learning depth-guided con- volutions for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition workshops, pages 1000–1001, 2020. 2

    work page 2020

  9. [9]

    Why does unsupervised pre-training help deep learning? In Proceedings of the thirteenth international con- ference on artificial intelligence and statistics , pages 201–

    Dumitru Erhan, Aaron Courville, Yoshua Bengio, and Pas- cal Vincent. Why does unsupervised pre-training help deep learning? In Proceedings of the thirteenth international con- ference on artificial intelligence and statistics , pages 201–

    work page

  10. [10]

    Fast-slow test- time adaptation for online vision-and-language navigation

    Junyu Gao, Xuan Yao, and Changsheng Xu. Fast-slow test- time adaptation for online vision-and-language navigation. In Forty-first International Conference on Machine Learn- ing, 2024. 2

    work page 2024

  11. [11]

    Are we ready for autonomous driving? the kitti vision bench- mark suite

    Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision bench- mark suite. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 3354–

    work page

  12. [12]

    Test time adaptation via conjugate pseudo-labels.Ad- vances in Neural Information Processing Systems, 35:6204– 6218, 2022

    Sachin Goyal, Mingjie Sun, Aditi Raghunathan, and J Zico Kolter. Test time adaptation via conjugate pseudo-labels.Ad- vances in Neural Information Processing Systems, 35:6204– 6218, 2022. 4

    work page 2022

  13. [13]

    Clust3: Information invariant test- time training

    Gustavo A Vargas Hakim, David Osowiechi, Mehrdad Noori, Milad Cheraghalikhani, Ali Bahri, Ismail Ben Ayed, and Christian Desrosiers. Clust3: Information invariant test- time training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6136–6145, 2023. 2

    work page 2023

  14. [14]

    Benchmarking neu- ral network robustness to common corruptions and perturba- tions

    Dan Hendrycks and Thomas Dietterich. Benchmarking neu- ral network robustness to common corruptions and perturba- tions. In International Conference on Learning Representa- tions, 2018. 1, 6, 18

    work page 2018

  15. [15]

    Gradient-based uncertainty for monocular depth estimation

    Julia Hornauer and Vasileios Belagiannis. Gradient-based uncertainty for monocular depth estimation. In Proceedings of the European Conference on Computer Vision, pages 613–

    work page

  16. [16]

    Seva: Leverag- ing single-step ensemble of vicinal augmentations for test- time adaptation

    Zixuan Hu, Yichun Hu, and Ling-Yu Duan. Seva: Leverag- ing single-step ensemble of vicinal augmentations for test- time adaptation. arXiv preprint arXiv:2505.04087, 2025. 2

    Pith/arXiv arXiv 2025

  17. [17]

    Beyond entropy: Region confi- dence proxy for wild test-time adaptation

    Zixuan Hu, Yichun Hu, Xiaotong Li, Shixiang Tang, and Ling-Yu Duan. Beyond entropy: Region confi- dence proxy for wild test-time adaptation. arXiv preprint arXiv:2505.20704, 2025. 1, 2

    Pith/arXiv arXiv 2025

  18. [18]

    Monodtr: Monocular 3d object detection with depth-aware transformer

    Kuan-Chih Huang, Tsung-Han Wu, Hung-Ting Su, and Win- ston H Hsu. Monodtr: Monocular 3d object detection with depth-aware transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 4012–4021, 2022. 2

    work page 2022

  19. [19]

    Test-time classifier adjustment module for model-agnostic domain generaliza- tion

    Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generaliza- tion. Advances in Neural Information Processing Systems , 34:2427–2440, 2021. 1

    work page 2021

  20. [20]

    Design of an image edge detection filter using the sobel operator

    Nick Kanopoulos, Nagesh Vasanthavada, and Robert L Baker. Design of an image edge detection filter using the sobel operator. IEEE Journal of solid-state circuits , 23(2): 358–367, 1988. 5

    work page 1988

  21. [21]

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 14162–14171, 2024. 2

    work page 2024

  22. [22]

    Entropy is not enough for test-time adaptation: From the perspective of disentangled factors

    Jonghyun Lee, Dahuin Jung, Saehyung Lee, Junsung Park, Juhyeon Shin, Uiwon Hwang, and Sungroh Yoon. Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. In The Twelfth International Confer- ence on Learning Representations, 2024. 2, 3, 6, 17

    work page 2024

  23. [23]

    A comprehensive sur- vey on test-time adaptation under distribution shifts

    Jian Liang, Ran He, and Tieniu Tan. A comprehensive sur- vey on test-time adaptation under distribution shifts. Inter- national Journal of Computer Vision, pages 1–34, 2024. 1

    work page 2024

  24. [24]

    Monotta: Fully test-time adaptation for monoc- ular 3d object detection

    Hongbin Lin, Yifan Zhang, Shuaicheng Niu, Shuguang Cui, and Zhen Li. Monotta: Fully test-time adaptation for monoc- ular 3d object detection. In Proceedings of the European Conference on Computer Vision , pages 96–114. Springer,

    work page

  25. [25]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Doll´ar. Focal loss for dense object detection. In Pro- ceedings of the IEEE international conference on computer vision, pages 2980–2988, 2017. 2, 3, 4

    work page 2017

  26. [26]

    Video test-time adaptation for action recognition

    Wei Lin, Muhammad Jehanzeb Mirza, Mateusz Kozinski, Horst Possegger, Hilde Kuehne, and Horst Bischof. Video test-time adaptation for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 22952–22961, 2023. 2

    work page 2023

  27. [27]

    Depth-aware test-time training for zero-shot video object segmentation

    Weihuang Liu, Xi Shen, Haolun Li, Xiuli Bi, Bo Liu, Chi- Man Pun, and Xiaodong Cun. Depth-aware test-time training for zero-shot video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19218–19227, 2024. 2

    work page 2024

  28. [28]

    Autoshape: Real-time shape-aware monoc- ular 3d object detection

    Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monoc- ular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15641– 15650, 2021. 2

    work page 2021

  29. [29]

    Bevfusion: Multi- task multi-sensor fusion with unified bird’s-eye view repre- sentation

    Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi- task multi-sensor fusion with unified bird’s-eye view repre- sentation. In 2023 IEEE international conference on robotics and automation (ICRA) , pages 2774–2781. IEEE, 2023. 6, 18

    work page 2023

  30. [30]

    Geometry uncer- tainty projection network for monocular 3d object detection

    Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncer- tainty projection network for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3111–3121, 2021. 1

    work page 2021

  31. [31]

    Latr: 3d lane detection from monocular images with transformer

    Yueru Luo, Chaoda Zheng, Xu Yan, Tang Kun, Chao Zheng, Shuguang Cui, and Zhen Li. Latr: 3d lane detection from monocular images with transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 7941–7952, 2023. 1

    work page 2023

  32. [32]

    Delving into localization errors for monocular 3d object detection

    Xinzhu Ma, Yinmin Zhang, Dan Xu, Dongzhan Zhou, Shuai Yi, Haojie Li, and Wanli Ouyang. Delving into localization errors for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4721–4730, 2021. 2, 3, 17

    work page 2021

  33. [33]

    3d object detection from images for autonomous driv- ing: a survey

    Xinzhu Ma, Wanli Ouyang, Andrea Simonelli, and Elisa Ricci. 3d object detection from images for autonomous driv- ing: a survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1

    work page 2023

  34. [34]

    Vision-centric bev perception: A survey

    Yuexin Ma, Tai Wang, Xuyang Bai, Huitong Yang, Yuenan Hou, Yaming Wang, Yu Qiao, Ruigang Yang, and Xinge Zhu. Vision-centric bev perception: A survey. IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 2024. 1

    work page 2024

  35. [35]

    Power-law escape rate of sgd

    Takashi Mori, Liu Ziyin, Kangqiao Liu, and Masahito Ueda. Power-law escape rate of sgd. In International Conference on Machine Learning, pages 15959–15975. PMLR, 2022. 4

    work page 2022

  36. [36]

    Efficient test-time model adaptation without forgetting

    Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In Interna- tional conference on machine learning, pages 16888–16905. PMLR, 2022. 6, 17

    work page 2022

  37. [37]

    Towards stable test-time adaptation in dynamic wild world

    Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In Internetional Conference on Learning Representations, 2023. 2, 3

    work page 2023

  38. [38]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in neural information processing systems, 2019. 6

    work page 2019

  39. [39]

    On the uncertainty of self-supervised monocular depth estimation

    Matteo Poggi, Filippo Aleotti, Fabio Tosi, and Stefano Mat- toccia. On the uncertainty of self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 3227–3237, 2020. 1, 5

    work page 2020

  40. [40]

    Monoground: Detecting monocular 3d objects from the ground

    Zequn Qin and Xi Li. Monoground: Detecting monocular 3d objects from the ground. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 3793–3802, 2022. 2, 6, 17

    work page 2022

  41. [41]

    Categorical depth distribution network for monocular 3d object detection

    Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555–8564, 2021. 2

    work page 2021

  42. [42]

    Uncertainty reduction for model adaptation in semantic segmentation

    Prabhu Teja S and Francois Fleuret. Uncertainty reduction for model adaptation in semantic segmentation. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9613–9623, 2021. 1

    work page 2021

  43. [43]

    Disen- tangling monocular 3d object detection

    Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel L ´opez-Antequera, and Peter Kontschieder. Disen- tangling monocular 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1991–1999, 2019. 7

    work page 1991

  44. [44]

    Test-time training with self- supervision for generalization under distribution shifts

    Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self- supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229–

  45. [45]

    Legendre-fenchel transforms in a nutshell

    Hugo Touchette. Legendre-fenchel transforms in a nutshell. URL http://www. maths. qmul. ac. uk/ht/archive/lfth2. pdf , page 25, 2005. 2, 4

    work page 2005

  46. [46]

    Tent: Fully test-time adaptation by entropy minimization

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Ol- shausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2020. 1, 3, 17

    work page 2020

  47. [47]

    Depth- conditioned dynamic message propagation for monocular 3d object detection

    Li Wang, Liang Du, Xiaoqing Ye, Yanwei Fu, Guodong Guo, Xiangyang Xue, Jianfeng Feng, and Li Zhang. Depth- conditioned dynamic message propagation for monocular 3d object detection. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , pages 454–463, 2021. 2

    work page 2021

  48. [48]

    Probabilistic and geometric depth: Detecting objects in per- spective

    Tai Wang, ZHU Xinge, Jiangmiao Pang, and Dahua Lin. Probabilistic and geometric depth: Detecting objects in per- spective. In Conference on Robot Learning , pages 1475–

    work page

  49. [49]

    Train in germany, test in the usa: Making 3d ob- ject detectors generalize

    Yan Wang, Xiangyu Chen, Yurong You, Li Erran Li, Bharath Hariharan, Mark Campbell, Kilian Q Weinberger, and Wei- Lun Chao. Train in germany, test in the usa: Making 3d ob- ject detectors generalize. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 11713–11723, 2020. 1

    work page 2020

  50. [50]

    Uncertainty-aware pseudo label refinery for domain adaptive semantic segmentation

    Yuxi Wang, Junran Peng, and ZhaoXiang Zhang. Uncertainty-aware pseudo label refinery for domain adaptive semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 9092–9101, 2021. 1

    work page 2021

  51. [51]

    Backpropagation-free network for 3d test-time adapta- tion

    Yanshuo Wang, Ali Cheraghian, Zeeshan Hayder, Jie Hong, Sameera Ramasinghe, Shafin Rahman, David Ahmedt- Aristizabal, Xuesong Li, Lars Petersson, and Mehrtash Ha- randi. Backpropagation-free network for 3d test-time adapta- tion. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 23231–23241,

    work page

  52. [52]

    Mononerd: Nerf- like representations for monocular 3d object detection

    Junkai Xu, Liang Peng, Haoran Cheng, Hao Li, Wei Qian, Ke Li, Wenxiao Wang, and Deng Cai. Mononerd: Nerf- like representations for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6814–6824, 2023. 1

    work page 2023

  53. [53]

    Monocd: Monocular 3d object detection with complementary depths

    Longfei Yan, Pei Yan, Shengzhou Xiong, Xuanyu Xiang, and Yihua Tan. Monocd: Monocular 3d object detection with complementary depths. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 10248–10257, 2024. 2, 17

    work page 2024

  54. [54]

    Deep layer aggregation

    Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 2403–2412, 2018. 17

    work page 2018

  55. [55]

    Objects are differ- ent: Flexible monocular 3d object detection

    Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are differ- ent: Flexible monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 3289–3298, 2021. 2, 6, 17

    work page 2021

  56. [56]

    DELTA: DEGRADATION-FREE FULLY TEST-TIME ADAPTA- TION

    Bowen Zhao, Chen Chen, and Shu-Tao Xia. DELTA: DEGRADATION-FREE FULLY TEST-TIME ADAPTA- TION. In The Eleventh International Conference on Learn- ing Representations, 2023. 2 Adaptive Dual Uncertainty Optimization: Boosting Monocular 3D Object Detection under Test-Time Shifts ————Appendix———— The structure of Appendix is as follows: • Appendix A contains a...

    work page arXiv 2023

  57. [58]

    Gaussian Noise Shot Noise Impulse Noise Original Image Defocus Glass Blur Motion Blur Snow Frost Fog Brightness Contrast Pixelate Saturate Figure 8

    Each corruption is further divided into five severity levels, with higher levels indicating more extreme perturbations and distribution shifts. Gaussian Noise Shot Noise Impulse Noise Original Image Defocus Glass Blur Motion Blur Snow Frost Fog Brightness Contrast Pixelate Saturate Figure 8. An illustration of 13 distinct types of corruptions in the sever...

    work page

  58. [208]

    JMLR Workshop and Conference Proceedings, 2010. 4

    work page 2010

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.