Pith. sign in

REVIEW 3 major objections 5 minor 61 references

Orthogonal Knowledge Refreshing for Domain-Incremental Object Detection

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Domain-incremental object detection can be made forgetting-free by expanding one low-rank branch per domain and projecting new gradients onto the orthogonal complement of an estimated historical subspace.

desk verdict Strong DIOD results, but the orthogonal projection in Eq. (6) is built from current-domain features, so the central mechanism is unsupported as written. read the letter →

arxiv 2607.17340 v1 pith:5CTMCXCO submitted 2026-07-19 cs.CV

Aoting Zhang , Dongbao Yang , Chang Liu , Xiaopeng Hong , Can Ma , Yu Zhou This is my paper
classification cs.CV
keywords domain-incrementalobjectdetectioncatastrophicforgettingparameter-efficientfine-tuninglow-rankadaptation(LoRA)orthogonalgradientprojectionexemplar-freecontinuallearningsubspaceapproximationtopology-awareconsistency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that catastrophic forgetting in domain-incremental object detection can be avoided without storing old images or identifying the current domain at inference. The recipe is to freeze the pretrained detector, add one small low-rank (LoRA) branch per new domain, fuse all branches into a single weight matrix, and during training project each new branch's gradient onto the orthogonal complement of an estimated historical subspace. A secondary regularizer aligns class prototypes across domains so that separately learned subspaces do not fragment shared semantic structure. If the approach is right, a detector can keep adapting to new styles, scenes, and corruptions while preserving old-domain accuracy, and the paper reports substantial gains over prior exemplar-free methods on three benchmark sequences.

What carries the argument

The load-bearing object is the historical subspace M_t, built by SVD of the new domain's feature matrix through the previously fused LoRA weights. Its work is to approximate the gradient subspace of past tasks; the update rule (6) then subtracts any component of a new gradient that lies in M_t. The second piece is low-rank subspace expansion: one LoRA per domain with linear fusion W_t = W_{t-1} + B_t A_t, giving fixed-parameter, no-routing inference.

What would settle it

Train on two domains in both orders. For each order, compute M_t from the new domain (as in the paper) and also from a held-out sample of the first domain; measure the principal angle between the two subspaces. If the angle is large and final mAP on the first domain is still high, the protection cannot be coming from Eq. (6) as written. Conversely, if mAP on the first domain drops when the angle is large, the subspace substitution is falsified as the mechanism.

Watch

Extended reading notes

Core claim

The central discovery is that domain-specific LoRA branches plus a gradient-projection rule yield conflict-free continual adaptation. At session t, the effective weight is W_t = W_0 + sum_{i<=t} B_i A_i; only the newest A_t, B_t are trained. Before training, the method feeds samples of the new domain through the frozen fused weights, collects per-layer feature matrices R_t, takes their SVD, and defines the historical subspace M_t as the span of the top-k left singular vectors (Eqs. 2-5). Every gradient step for the new branch is then corrected by Eq. (6): ∇ = ∇ - ∇ M_t M_t^T, so the update lies in the orthogonal complement of M_t. The paper claims this suppresses interference with past knowl

Load-bearing premise

The method assumes that the feature matrix of the current domain, passed through the fused previous weights, spans the gradient directions that mattered for earlier domains; if that span is wrong, the orthogonal projection in Eq. (6) protects no old knowledge and only removes useful update directions.

Editorial extensions

If this is right

  • A detector can be adapted across many domains with a fixed parameter budget, because LoRA branches fuse additively and no domain selector is needed at inference.
  • Exemplar-free domain-incremental detection does not require replay: orthogonal constraints on new branches are sufficient to retain old-domain accuracy, with reported final-session gains of +5.6 mAP on Pascal VOC and +6.5 mAP on BDD100K over the best prior exemplar-free method.
  • Projecting gradients into unused directions provides both stability and plasticity, letting the method approach the upper-bound performance on the BDD100K sequence.
  • The topology-aware consistency loss keeps class prototypes in new domains aligned with base-domain prototypes, preserving cross-domain semantic structure across incremental sessions.
  • The method also handles long sequences of low-level corruptions (16 domains in 5 sessions), maintaining a better stability-plasticity balance than bias-based and prompt-based alternatives.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The orthogonal-refreshing recipe is not tied to object detection: any parameter-efficient fine-tuning scheme whose updates live in a low-rank space could use the same projection to protect previous tasks, so the core mechanism may transfer to other continual-learning settings such as segmentation or classification.
  • The weakest link is the substitution of the new domain's feature span for the old domains' gradient subspace. A direct test would be to compute M_t from a held-out set of old-domain features and compare it with the paper's M_t; if the two differ substantially, the reported gains are likely carried by the LoRA expansion and topology loss rather than by the orthogonality projection.
  • The method implies order-dependent behavior: if the historical subspace is estimated from whichever domain arrives second, swapping the order of two domains should change what is protected. Testing order invariance would separate the approximation's effect from the expansion's effect.
  • The prototype alignment assumes a shared class set across domains; if the label space itself drifts, the topology-aware consistency term would need a mechanism for inserting new classes, which the paper does not address.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes OKR, an exemplar-free framework for domain-incremental object detection. OKR injects a dedicated LoRA branch for each new domain into a frozen ViT backbone, fuses all branches via the linear additivity of LoRA, and constrains new-domain gradient updates to be orthogonal to an estimated 'historical subspace' (Eq. 6). This subspace is obtained by SVD of the feature matrix of the current domain, computed under the old fused weights W_{t-1} (Eqs. 2-5). In addition, a topology-aware consistency loss aligns class prototypes across domains. Experiments on Pascal VOC, BDD100K, and VOC-Corruption report large gains over prior exemplar-free methods, with ablations attributing part of the gain to the orthogonal refreshing strategy.

Significance. If the mechanism were sound, OKR would be a useful contribution to exemplar-free DIOD: it combines parameter-efficient adaptation with a simple gradient-orthogonality constraint and reports substantial mAP improvements over strong baselines. The empirical study is broad (three benchmarks, multiple ablations, efficiency analysis) and the writing is mostly clear. However, the central theoretical justification for the orthogonal refreshing step is not valid as stated: the subspace M_t is built from current-domain features rather than historical-task features, and the SVD/projection in Eq. (6) is dimensionally ill-defined. The empirical gains do not by themselves establish the claimed 'historical orthogonality' mechanism, so the paper's core claim is currently unsupported.

major comments (3)
  1. [§4.3, Eqs. (2)–(6)] The subspace M_t is computed from the current domain D_t under old fused weights W_{t-1} (Eq. (2)), not from previous tasks' inputs. The cited result [30] states that gradient updates of a linear layer lie in the span of that layer's inputs for the task being updated. To protect prior knowledge, the relevant span is that of previous tasks' inputs. No argument is given that current-domain features under W_{t-1} span the historical gradient directions. In fact, by the same cited property, the current gradient is dominated by current inputs, so Eq. (6) removes much of the signal needed to fit D_t. The ablation (Exp. #4 vs #3, Table 5) shows only an empirical benefit on the benchmark; it does not validate the historical-orthogonality mechanism. This is a load-bearing gap because GOR is the core novelty.
  2. [§4.3, Eqs. (3) and (6)] The SVD in Eq. (3) is dimensionally inconsistent with the projection in Eq. (6). If R_t ∈ R^{m×n} with m samples and n feature dimensions, then U_t ∈ R^{m×m} and M_t M_t^T ∈ R^{m×m}. Right-multiplying the gradient ∇_w L_t, where w = {A_t, B_t} are LoRA parameters of shape r×d_in and d_out×r, by an m×m matrix is generally undefined. GPM builds a projection in the feature space using singular vectors with the appropriate orientation; here no per-layer dimensional convention is given, so Eq. (6) is not a well-formed operation as written.
  3. [§4.3, Eqs. (1)–(2)] W_{t-1} is defined inconsistently. In Eq. (1), W_t = W_0 + Σ_{i=1}^t B_i A_i; the text after Eq. (2) defines W_{t-1} = Σ_{i=1}^{t-1} B_i A_i, omitting W_0. Since R_t is computed with the linear layer parameterized by W_{t-1}, this changes the feature matrix and hence the subspace. The paper must state exactly which matrix is used and justify the choice.
minor comments (5)
  1. [Fig. 1(c) and §1] The claim '+17.6% mAP gain on Comic dataset' does not specify the baseline. Table 2 suggests the comparison is to LDB (37.0 → 54.6), but the text and caption should be explicit.
  2. [§5.1 and Table 4] The text says the VOC-Corruption series includes 15 corrupted domains, but Table 4 lists 16 columns including Clean. Clarify whether 'domain' includes the clean base.
  3. [§5.2] The sentence 'OKR a superior balance of knowledge retention and adaptability' is missing a verb; should read 'OKR achieves a superior balance...'.
  4. [§4.2] The phrase 'without incurring parameter growth' is misleading because a new LoRA branch is added per domain. Clarify that the claim refers to inference after merging, or revise the wording.
  5. [References] Reference [30] (GPM) is cited as an arXiv preprint; a published version exists and should be cited.

Circularity Check

0 steps flagged · score 0.0 of 10

No constructional circularity: benchmark gains are against external methods and GOR is an explicit gradient transform; the D_t-based M_t is a validity gap, not a circular reduction.

full rationale

The paper's derivation chain does not reduce to its inputs. The +5.6%/+6.5% mAP claims are benchmark comparisons against external exemplar-free baselines (LDB, LDB+SOYO), not fitted predictions renamed as results. The gradient-based orthogonal refreshing is an explicit transformation (Eq. 6) of the current gradient using a basis obtained from SVD (Eqs. 2-5); although the basis M_t is computed from the current domain D_t rather than from old-task inputs, this is an unjustified approximation/validity concern about the claim that it represents a historical subspace, not an identity between the claimed preservation mechanism and the new-domain data. The cited gradient-span connection [30] is external prior work, and the self-citations ([48]-[51], [54]-[56]) occur in related work and are not load-bearing. No uniqueness theorem is imported from the authors, and no ansatz is smuggled via a self-citation. Hence, under the strict circularity standard, no step exhibits Eq. X = Eq. Y by construction or a fitted parameter called a prediction.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The largest unfunded premise is the identification of the historical gradient subspace with the current domain's feature span under the fused old weights. The method also depends on hand-chosen hyperparameters (rank r=16, SVD threshold ε, and an implicit TAC weight) that are not fully reported.

free parameters (3)
  • SVD energy threshold ε = not reported
    Eq. (4) uses ε to choose the number k of singular vectors; the paper gives no value or sensitivity analysis, and k determines the projection subspace M_t.
  • LoRA rank r = 16
    Ablation Fig. 5(a) shows performance plateaus beyond 16; 16 is a hand-chosen trade-off between capacity and efficiency.
  • TAC loss weight = 1 (implicit)
    Eq. (8) sums L_det and L_tac with no coefficient; no weighting or tuning is reported.
assumptions (5)
  • standard math Gradient of a linear layer lies in the span of its input features (from GPM).
    Invoked in §4.3 to justify approximating the gradient space by the feature matrix R_t and using SVD to construct M_t.
  • ad hoc to paper Current-domain features computed with fused W_{t-1} approximate the historical gradient subspace of previous tasks.
    Eqs. (2)–(5) use D_t, not historical data, to build M_t. No theoretical or empirical evidence is given that this substitution is valid.
  • domain assumption Freezing previous LoRA branches and fusing weights preserves prior knowledge; only gradient interference during fusion needs to be removed.
    Central design assumption in §4.2; not directly verified independently of the full method.
  • domain assumption Base-domain prototypes provide an invariant semantic topology for all later domains.
    §4.4 stores base prototypes and keeps them fixed; assumes the base domain is representative and that no class-level semantic shift occurs.
  • domain assumption ImageNet-1K pretrained ViT features have enough capacity for all target domains; low-rank updates suffice.
    Motivated by the oRecall diagnostic in §3.1, but only tested for the chosen backbones and domains.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Orthogonal Knowledge Refreshing for Domain-Incremental Object Detection." pith.science (2026). https://pith.science/paper/5CTMCXCO

@misc{pith2026260717340,
  author       = {Pith},
  title        = {Pith review of: Orthogonal Knowledge Refreshing for Domain-Incremental Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5CTMCXCO}},
  note         = {Machine review of arXiv:2607.17340}
}
read the original abstract

Domain-incremental object detection (DIOD) requires models to continually adapt to new domains while preserving prior knowledge. Recently, parameter-efficient fine-tuning offers a promising avenue, wherein a pre-trained model is frozen and a small number of learnable parameters are injected for downstream tasks. However, these methods risk overwriting critical past knowledge, triggering inter-domain interference and performance degradation. To address this challenge, we propose Orthogonal Knowledge Refreshing (OKR), a simple yet effective framework for DIOD. OKR incrementally constructs independent domain-specific subspaces via dedicated low-rank branches for each domain, which are seamlessly fused for a holistic decision, enabling conflict-free capacity expansion without domain selection during inference. To minimize knowledge interference during fusion, we present a gradient-based orthogonal refreshing strategy that projects gradient updates of new domains onto the orthogonal complement of the fused historical subspace, supporting continual adaptation without forgetting. Moreover, to mitigate semantic fragmentation across domains, we enforce topology-aware consistency, aligning the semantic structures of old and new domains. Extensive experiments validate the superiority of OKR, outperforming the best exemplar-free method by significant margins of +5.6% and +6.5% mAP on the Pascal VOC and BDD100K series, respectively.

Figures

Figures reproduced from arXiv: 2607.17340 by the authors.

Figure 1
Figure 1. (a) Previous methods suffer from inter-domain interference due to shared pa￾rameter overwriting. (b) OKR enables conflict-free adaptation by expanding dedicated LoRA branches for each session. Through orthogonal knowledge refreshing, it projects gradients orthogonally to the fused historical subspace, preserving prior knowledge while bypassing domain routing. (c) OKR achieves superior performance, delivering a signi… view at source ↗
Figure 2
Figure 2. Class-agnostic object recall (oRecall) on different domains over learning ses￾sions. Solid lines indicate the session in which each domain is learned. work addresses continual adaptation across sequential domain shifts without forgetting prior domains. 3 Preliminary 3.1 Observations Performance degradation under domain shift primarily stems from feature distri￾bution misalignment [25, 40, 47]. While mapping source a… view at source ↗
Figure 3
Figure 3. Overview of OKR. To expand model capacity without conflicts, OKR incremen￾tally constructs independent domain-specific subspaces by injecting dedicated LoRAs into the feature extractor. At session t, only the new LoRA (purple) is trainable, while previously ones (blue) remain frozen. During training and inference, all branches are seamlessly fused for holistic decision-making without redundant forward passes. To fur… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Performance across sessions on VOC and BDD100K series. OKR consis￾tently outperforms PEFT-based methods [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Further analysis of OKR. (a) OKR remains robust under different ranks of the expandable LoRA subspaces. (b) Topology-aware consistency preserves semantic topology across domain increments. (c) GOR maintains stable optimization, with sim￾ilar training loss curves with a…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 6 linked inside Pith

  1. [30]

    arXiv preprint arXiv:2103.09762 (2021)

    Saha, G., Garg, I., Roy, K.: Gradient projection memory for continual learning. arXiv preprint arXiv:2103.09762 (2021)

    arXiv 2021

  2. [1]

    In: CVPR

    Bang, J., Kim, H., Yoo, Y., Ha, J.W., Choi, J.: Rainbow Memory: Continual learn- ing with a memory of diverse samples. In: CVPR. pp. 8218–8227 (2021)

    2021

  3. [2]

    arXiv preprint arXiv:2505.15649 (2025)

    Cao, T., Lyu, J., Zeng, W., Mu, W., Zhou, Y.: The devil is in fine-tuning and long-tailed problems: a new benchmark for scene text detection. arXiv preprint arXiv:2505.15649 (2025)

    arXiv 2025

  4. [3]

    arXiv preprint arXiv:1812.00420 (2018)

    Chaudhry, A., Ranzato, M., Rohrbach, M., Elhoseiny, M.: Efficient lifelong learning with a-gem. arXiv preprint arXiv:1812.00420 (2018)

    arXiv 2018

  5. [4]

    In: CVPR

    Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR. pp. 3213–3223 (2016)

    2016

  6. [5]

    In: IEEE TCSVT

    Ding, L., Song, X., He, Y., Wang, C., Dong, S., Wei, X., Gong, Y.: Domain incre- mental object detection based on feature space topology preserving strategy. In: IEEE TCSVT. vol. 34, pp. 424–437 (2023)

    2023

  7. [6]

    In: CVPR

    Ding, N., Xu, Y., Tang, Y., Xu, C., Wang, Y., Tao, D.: Source-free domain adap- tation via distribution estimation. In: CVPR. pp. 7212–7222 (2022)

    2022

  8. [7]

    In: CVPR

    Douillard,A.,Ramé,A.,Couairon,G.,Cord,M.:Dytox:Transformersforcontinual learning with dynamic token expansion. In: CVPR. pp. 9285–9295 (2022)

    2022

Show all 61 references
  1. [8]

    Du, Y., Yang, H., Chen, M., Luo, H., Jiang, J., Xin, Y., Wang, C.: Generation, augmentation, and alignment: A pseudo-source domain based method for source- free domain adaptation. In: ML. vol. 113, pp. 3611–3631 (2024)

    2024

  2. [9]

    In: IJCV

    Everingham,M.,VanGool,L.,Williams,C.K.,Winn,J.,Zisserman,A.:Thepascal visual object classes (voc) challenge. In: IJCV. vol. 88, pp. 303–338 (2010) 16 A. Zhang et al

    2010

  3. [10]

    In: CVPR

    Feng, T., Wang, M., Yuan, H.: Overcoming catastrophic forgetting in incremental object detection via elastic response distillation. In: CVPR. pp. 9427–9436 (2022)

    2022

  4. [11]

    French, R.M.: Catastrophic forgetting in connectionist networks. In: TCS. vol. 3, pp. 128–135 (1999)

    1999

  5. [12]

    In: ICML

    Gao, R., Liu, W.: DDGR: Continual learning with deep diffusion-based generative replay. In: ICML. pp. 10744–10763 (2023)

    2023

  6. [13]

    In: ICME

    Hao, Y., Fu, Y., Jiang, Y.G., Tian, Q.: An end-to-end architecture for class- incremental object detection with knowledge distillation. In: ICME. pp. 1–6 (2019)

    2019

  7. [14]

    arXiv preprint arXiv:1903.12261 (2019)

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

    1903 arXiv

  8. [15]

    In: CVPR

    Hu, X., Fu, C.W., Zhu, L., Heng, P.A.: Depth-attentional features for single-image rain removal. In: CVPR. pp. 8022–8031 (2019)

    2019

  9. [16]

    In: NeurIPS

    Huang, J., Guan, D., Xiao, A., Lu, S.: Model Adaptation: Historical contrastive learning for unsupervised domain adaptation without source data. In: NeurIPS. vol. 34, pp. 3635–3649 (2021)

    2021

  10. [17]

    In: AAAI

    Huang, L., Zeng, Y., Yang, C., An, Z., Diao, B., Xu, Y.: eTag: Class-incremental learning via embedding distillation and task-oriented generation. In: AAAI. vol. 38, pp. 12591–12599 (2024)

    2024

  11. [18]

    In: CVPR

    Inoue, N., Furuta, R., Yamasaki, T., Aizawa, K.: Cross-domain weakly-supervised object detection through progressive domain adaptation. In: CVPR. pp. 5001–5009 (2018)

    2018

  12. [19]

    In: ECCV

    Jin, Y., Wang, X., Long, M., Wang, J.: Minimum class confusion for versatile domain adaptation. In: ECCV. pp. 464–480 (2020)

    2020

  13. [20]

    IEEE TPAMI45(7), 9022–9040 (2023)

    Li,W.,Liu,X.,Yuan,Y.:SIGMA++:improvedsemantic-completegraphmatching for domain adaptive object detection. IEEE TPAMI45(7), 9022–9040 (2023)

    2023

  14. [21]

    In: ECCV

    Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: ECCV. pp. 280–296 (2022)

    2022

  15. [22]

    In: IEEE TPAMI

    Li, Z., Hoiem, D.: Learning without forgetting. In: IEEE TPAMI. vol. 40, pp. 2935–2947 (2017)

    2017

  16. [23]

    In: ICML

    Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: ICML. pp. 6028–6039 (2020)

    2020

  17. [24]

    In: CVPR

    Liang, Y.S., Li, W.J.: Adaptive plasticity improvement for continual learning. In: CVPR. pp. 7816–7825 (2023)

    2023

  18. [25]

    In: ICML

    Long, M., Cao, Y., Wang, J., Jordan, M.: Learning transferable features with deep adaptation networks. In: ICML. vol. 37, pp. 97–105 (July 2015)

    2015

  19. [26]

    In: CVPR

    Lu, Y., Liu, J., Zhang, Y., Liu, Y., Tian, X.: Prompt distribution learning. In: CVPR. pp. 5206–5215 (June 2022)

    2022

  20. [27]

    In: CVIU

    Peng, C., Zhao, K., Maksoud, S., Li, M., Lovell, B.C.: SID: Incremental learning for anchor-free object detection via selective and inter-related distillation. In: CVIU. vol. 210, p. 103229 (2021)

    2021

  21. [28]

    In: NeurIPS

    Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. In: NeurIPS. vol. 28, pp. 91–99 (2015)

    2015

  22. [29]

    In: NeurIPS

    Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T., Wayne, G.: Experience replay for continual learning. In: NeurIPS. vol. 32, pp. 350–360 (2019)

    2019

  23. [31]

    In: CVPR (June 2019)

    Saito, K., Ushiku, Y., Harada, T., Saenko, K.: Strong-weak distribution alignment for adaptive object detection. In: CVPR (June 2019)

    2019

  24. [32]

    In: NeurIPS

    Shi, H., Wang, H.: A unified approach to domain incremental learning with mem- ory: Theory and algorithm. In: NeurIPS. vol. 36, pp. 15027–15059 (2023) Orthogonal Knowledge Refreshing 17

    2023

  25. [33]

    In: AAAI

    Song, X., He, Y., Dong, S., Gong, Y.: Non-exemplar domain incremental object detection via learning domain bias. In: AAAI. vol. 38, pp. 15056–15065 (2024)

    2024

  26. [34]

    In: CVPR

    Tang,S.,Chen,D.,Zhu,J.,Yu,S.,Ouyang,W.:Layerwiseoptimizationbygradient decomposition for continual learning. In: CVPR. pp. 9634–9643 (2021)

    2021

  27. [35]

    arXiv preprint arXiv:2107.12585 (2021)

    Tang, S., Yang, Y., Ma, Z., Hendrich, N., Zeng, F., Ge, S.S., Zhang, C., Zhang, J.: Nearest neighborhood-based deep clustering for source data-absent unsupervised domain adaptation. arXiv preprint arXiv:2107.12585 (2021)

    2021 arXiv

  28. [36]

    In: IEEE TCSVT

    Tian, J., Zhang, J., Li, W., Xu, D.: VDM-DA: Virtual domain modeling for source data-free domain adaptation. In: IEEE TCSVT. vol. 32, pp. 3749–3760 (2021)

    2021

  29. [37]

    In: CVPR

    VS, V., Gupta, V., Oza, P., Sindagi, V.A., Patel, V.M.: MeGA-CDA: Memory guided attention for category-aware unsupervised domain adaptive object detec- tion. In: CVPR. pp. 4516–4526 (June 2021)

    2021

  30. [38]

    In: CVPR

    VS, V., Oza, P., Patel, V.M.: Instance relation graph guided source-free domain adaptive object detection. In: CVPR. pp. 3520–3530 (2023)

    2023

  31. [39]

    In: ICLR (2022)

    Wang, F.Y., Zhou, D.W., Liu, L., Ye, H.J., Bian, Y., Zhan, D.C., Zhao, P.: BEEF: Bi-compatible class-incremental learning via energy-based expansion and fusion. In: ICLR (2022)

    2022

  32. [40]

    PR137, 109319 (2023)

    Wang, M., Wang, S., Wang, W., Shen, L., Zhang, X., Lan, L., Luo, Z.: Reducing bi-level feature redundancy for unsupervised domain adaptation. PR137, 109319 (2023)

    2023

  33. [41]

    In: CVPR

    Wang, Q., Song, X., He, Y., Han, J., Ding, C., Gao, X., Gong, Y.: Boosting domain incremental learning: Selecting the optimal parameters is all you need. In: CVPR. pp. 4839–4849 (June 2025)

    2025

  34. [42]

    arXiv preprint arXiv:2310.14152 (2023)

    Wang, X., Chen, T., Ge, Q., Xia, H., Bao, R., Zheng, R., Zhang, Q., Gui, T., Huang, X.: Orthogonal subspace learning for language model continual learning. arXiv preprint arXiv:2310.14152 (2023)

    2023 arXiv

  35. [43]

    In: NeurIPS

    Wang, Y., Huang, Z., Hong, X.: S-Prompts learning with pre-trained transformers: An occam’s razor for domain incremental learning. In: NeurIPS. vol. 35, pp. 5682– 5695 (2022)

    2022

  36. [44]

    In: NeurIPS

    Wang, Y., Chauhan, J., Wang, W., Hsieh, C.J.: Universality and limitations of prompt tuning. In: NeurIPS. vol. 36, pp. 75623–75643 (2023)

    2023

  37. [45]

    In: ECCV

    Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: DualPrompt: Complementary prompting for rehearsal-free continual learning. In: ECCV. pp. 631–648 (2022)

    2022

  38. [46]

    In: CVPR

    Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: CVPR. pp. 139–149 (2022)

    2022

  39. [47]

    In: AAAI

    Xie, B., Yuan, L., Li, S., Liu, C.H., Cheng, X., Wang, G.: Active learning for domain adaptation: An energy-based approach. In: AAAI. vol. 36, pp. 8708–8716 (2022)

    2022

  40. [48]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Yang, D., Zhou, Y., Hong, X., Zhang, A., Wang, W.: One-shot replay: Boosting incremental object detection via retrospecting one object. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 3127–3135 (2023)

    2023

  41. [49]

    In: Proceedings of the 31st ACM International Conference on Multimedia

    Yang, D., Zhou, Y., Hong, X., Zhang, A., Wei, X., Zeng, L., Qiao, Z., Wang, W.: Pseudo object replay and mining for incremental object detection. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 153–162 (2023)

    2023

  42. [50]

    ACM Transac- tions on Multimedia Computing, Communications, and Applications18(1), 1–23 (2022) 18 A

    Yang, D., Zhou, Y., Shi, W., Wu, D., Wang, W.: RD-IOD: Two-level residual- distillation-based triple-network for incremental object detection. ACM Transac- tions on Multimedia Computing, Communications, and Applications18(1), 1–23 (2022) 18 A. Zhang et al

    2022

  43. [51]

    Yang, D., Zhou, Y., Zhang, A., Sun, X., Wu, D., Wang, W., Ye, Q.: Multi-view correlation distillation for incremental object detection. In: PR. vol. 131, p. 108863 (2022)

    2022

  44. [52]

    In: ICCV

    Yao, X., Zhao, S., Xu, P., Yang, J.: Multi-source domain adaptation for object detection. In: ICCV. pp. 3273–3282 (2021)

    2021

  45. [53]

    In: CVPR

    Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In: CVPR. pp. 2636–2645 (2020)

    2020

  46. [54]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Zhang, A., Yang, D., Liu, C., Hong, X., Shang, M., Zhou, Y.: DCA: Dividing and conquering amnesia in incremental object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 9851–9859 (2025)

    2025

  47. [55]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Zhang, A., Yang, D., Liu, C., Hong, X., Zhou, Y.: Specifying what you know or not for multi-label class-incremental learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 22345–22353 (2025)

    2025

  48. [56]

    arXiv preprint arXiv:2606.15253 (2026)

    Zhang, A., Yang, D., Liu, C., Hong, X., Zhou, Y.: Focus, Align, and Sustain: Counteracting gradient dilution in incremental object detection. arXiv preprint arXiv:2606.15253 (2026)

    2026

  49. [57]

    In: International Conference on Document Analysis and Recognition

    Zhang, D., Lyu, J., Shen, Z., Zhou, Y.: Class-agnostic region-of-interest match- ing in document images. In: International Conference on Document Analysis and Recognition. pp. 446–464 (2025)

    2025

  50. [58]

    Visual Intelligence4(1), 15 (2026)

    Zhang, L., Qian, L., Li, R., Li, T., Zhang, W.: Leveraging multi-modal and histor- ical knowledge graphs for continual robot navigation. Visual Intelligence4(1), 15 (2026)

    2026

  51. [59]

    IEEE TPAMI (2024)

    Zhou,D.W.,Wang,Q.W.,Qi,Z.H.,Ye,H.J.,Zhan,D.C.,Liu,Z.:Class-incremental learning: A survey. IEEE TPAMI (2024)

    2024

  52. [60]

    In: CVPR

    Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision- language models. In: CVPR. pp. 16816–16825 (2022)

    2022

  53. [61]

    In: CVPR

    Zhu, F., Zhang, X.Y., Wang, C., Yin, F., Liu, C.L.: Prototype augmentation and self-supervision for incremental learning. In: CVPR. pp. 5871–5880 (2021)

    2021

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.