Pith. sign in

REVIEW 5 major objections 5 minor 61 references

iDPA: Instance Decoupled Prompt Attention for Incremental Medical Object Detection

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that incremental medical object detection succeeds on a frozen vision-language detector when prompts are generated from decoupled instance regions and injected through decoupled prompt attention.

desk verdict Useful benchmark and a plausible prompt-decoupling method, but the SOTA claim rests on baseline reproductions the paper itself admits may be unfair, and the theory section does not hold. read the letter →

arxiv 2506.00406 v1 pith:LE4MAGRM submitted 2025-05-31 cs.CV

Huahui Yi , Wei Xu , Ziyuan Qin , Xi Chen , Xiaohu Wu , Kang Li , Qicheng Lao This is my paper
classification cs.CV
keywords incrementalmedicalobjectdetectioncontinuallearningprompttuningvision-languagedecoupledattentioncatastrophicforgettingfew-shotimaging
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

This paper tries to establish that prompt-based continual learning can work for medical object detection, not just classification, once prompt knowledge is decoupled at two levels: instance-level prompts generated from object regions rather than whole images, and prompt attention decoupled from the long image-text token stream. It introduces the iDPA framework, built on a frozen pre-trained vision-language detection model, and validates it on ODinM-13, a new benchmark spanning 13 medical datasets, 8 imaging modalities, and 9 organs. The central claim is that iDPA outperforms existing state-of-the-art continual learning methods by 5.44% in final average precision under full data and by 4.83%, 12.88%, and 4.59% in the 1-shot, 10-shot, and 50-shot settings, while using only 1.4% trainable parameters and reducing catastrophic forgetting compared with prior prompt-based methods. A sympathetic reader would care because this points toward updating medical detectors incrementally as new lesion and organ categories appear, without storing old patient data or retraining full models.

What carries the argument

The load-bearing mechanism is Decoupled Prompt Attention (DPA), paired with Instance-level Prompt Generation (IPG). IPG uses RoI pooling to extract per-category instance representations from bounding-box regions and a cross-attention query-answer step called CCPKI to distill those representations into concept prompts, so the prompt encodes object-level knowledge rather than global image statistics. DPA rewrites the prompt attention of the vision-language fusion module so that the text feature update becomes $\bar{f}_t = \mathrm{Attn}_{v\to t}(f_v, f_t) + \lambda\, \mathrm{Attn}_{v\to t}(p_v, f_t)$: the original multimodal attention is preserved and prompt knowledge enters as a scaled, decoupled cross-attention term instead of being prepended to the token stream. This decoupled term is what carries instance knowledge into the frozen model; it also decouples prompt length from token length, which lowers memory and computation and leaves the original attention distribution intact.

What would settle it

A direct falsifier is to re-run iDPA on ODinM-13 with the dropped branch restored, keeping the prompt-to-prompt and feature-to-prompt attention terms alongside the decoupled branch, under identical training budgets; if final average precision rises by more than the reported run-to-run variance, the claimed lossless decoupling is false. A sharper version is a synthetic detection task where the background strongly correlates with the target class, since the dropped branch is where background-suppression information would live.

Watch

Extended reading notes

Core claim

The paper's central claim is that a frozen pre-trained vision-language object detector can be continually extended to new medical concepts if prompt knowledge is decoupled rather than prepended. Instance-level Prompt Generation (IPG) first extracts region-level features per category with RoI pooling and distills them into concept prompts through a cross-attention query-answer step, so prompts carry fine-grained object information instead of global foreground-background mixtures. Decoupled Prompt Attention (DPA) then rewrites the multimodal fusion attention so the enhanced text feature is the original vision-to-text attention plus a learnably scaled prompt-to-text term, dropping the coupled prompt-to-prompt and feature-to-prompt branches that dilute short prompts against very long token sequences. On the collected 13-dataset benchmark ODinM-13, this combination yields the best final average precision among compared continual learning methods in full-data and few-shot settings, with lower final forgetting and a small parameter footprint.

Load-bearing premise

The load-bearing premise is that decoupling is lossless: everything the full prompt attention could contribute to detection is carried by the single remaining branch of prompt-to-vision and vision-to-text knowledge injection, so dropping the prompt-to-prompt and feature-to-prompt branch and replacing its attention weight with a learnable scalar removes no useful information.

Editorial extensions

If this is right

  • If the central claim is correct, medical object detectors can be updated on new categories with only 1.4% of parameters trainable, without exemplar storage or retraining the backbone.
  • The reported 12.88% FAP gain in the 10-shot setting indicates that decoupled instance prompts are most valuable when annotations are scarce, the common clinical condition.
  • Prompt injection at the fusion encoder, rather than at the visual or text encoder, gives the best performance-cost balance, so future multimodal continual-learning designs should target the fusion stage.
  • The method also transfers to a domain-incremental polyp benchmark across different medical centers, suggesting the decoupling addresses domain shift as well as new categories.
  • The reported memory reduction relative to naive prompt attention means the benefit is not only accuracy but also lower training cost.

Reading between the lines

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

  • Beyond the paper, a testable extension of its reasoning is that DPA's advantage should grow with image resolution: the decoupled term removes the dilution of short prompts by very long token sequences, so high-resolution medical inputs should benefit more than low-resolution ones.
  • Beyond the paper, the lossless-decoupling assumption implies a falsifiable prediction: restoring the dropped prompt-to-prompt and feature-to-prompt branch after training should not change final average precision; a measured gain from reinserting that branch would contradict the derivation.
  • Beyond the paper, the same IPG-plus-DPA pattern could be applied to continual medical segmentation or other dense-prediction vision-language tasks, where foreground-background coupling is equally problematic.
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

5 major / 5 minor

Summary. The paper proposes iDPA, a prompt-based continual learning framework for incremental medical object detection built on a frozen GLIP backbone. It contributes two modules: Instance-level Prompt Generation (IPG), which extracts RoI-level instance features and uses cross-attention with residual scaling to generate task prompts, and Decoupled Prompt Attention (DPA), which modifies the cross-modal attention in the fusion encoder by dropping the prompt-to-prompt and feature-to-prompt branches and introducing a learnable scalar to scale the retained prompt-to-feature branch. The authors introduce a 13-dataset benchmark ODinM-13 and report FAP/CAP/FFP improvements over existing prompt-based methods in full-data, 1-shot, 10-shot, and 50-shot settings, with roughly 1.4% trainable parameters. An appendix contains a convergence analysis of DPA, implementation details, and additional tables.

Significance. If the empirical comparisons are fair and reproducible, iDPA would be a practically attractive method for incremental medical object detection: it reports large gains over NoRGa, CODA, and DIKI in both full-data and few-shot regimes, shows reduced forgetting, and the code and benchmark ODinM-13 are valuable community assets. The ablations on injection position, prompt length, and scaling factor are informative. However, the central SOTA claim is weakened by the paper's own admission that prompt-based baselines were prompted in all layers rather than in their original configurations, and the theoretical derivation of DPA is not a true equivalence. The empirical claim may still hold, but the current evidence does not yet establish it at the stated confidence.

major comments (5)
  1. [Appendix B] The baseline comparison is undermined by the reproduction protocol. Appendix B states: "We reproduce other prompt-based methods on GLIP by prompting all layers of both the vision and text encoders, whereas the original papers typically use only the embedding layer or a few initial layers (e.g., the first five layers). This discrepancy may lead to suboptimal performance on the IMOD task." Since the headline FAP gains in Tables 1 and 2 are computed against these reproduced baselines, the central claim that iDPA outperforms SOTA is not yet established. Please re-run all prompt-based baselines in their original prompt-placement configurations, report both configurations, and base the main comparison on the stronger baseline configuration.
  2. [Sec. 4.3, Eqs. (10)-(13)] The derivation from prompt attention (PA) to DPA is not an equivalence. Eq. (10) is the standard attention formula with prepended prompts; Eq. (12) then discards the [\bar{p}_t] branch and redefines the attention weight as a learnable scalar, and Eq. (13) presents the resulting expression as DPA. This is a method design choice, not a derivation, and dropping branch A may lose information needed for vision-text alignment. Please present DPA as an architectural modification motivated by efficiency and ablation evidence, rather than as an equivalent reformulation.
  3. [Appendix A.3, Lemma A.2] The convergence proof is circular. Lemma A.2 assumes "there be an optimal representation f* = B", where B is exactly the branch retained by DPA, and further assumes M1A = 0 at convergence, which states that the dropped branch contributes nothing to the output. Under these assumptions the conclusion that DPA performs as well as full prompt attention follows by construction and does not establish any efficiency or convergence benefit. Please remove this lemma or replace it with a non-circular analysis that does not assume the dropped branch is irrelevant.
  4. [Sec. 5.1 vs. Appendix B and Fig. 5c] The learning rate is reported inconsistently. Section 5.1 states "The learning rate is set to 0.1", while Fig. 5c shows that a grid search finds 0.01 optimal, and Appendix B states that prompt-based methods use an initial learning rate of 1e-2. Please specify exactly which learning rate was used for iDPA and for each baseline in each experiment, and correct the apparent contradiction, since this directly affects reproducibility.
  5. [Tables 1 and 2] No uncertainty measures are reported for the baseline methods. All results are averaged over only three seeds, and Tables 1 and 2 report single FAP/CAP/FFP values without standard deviations or confidence intervals. Given that the claimed margins are 4-5 AP points in several settings, seed-level noise could be a large fraction of the reported gap. Please report per-seed results or error bars for all methods, or provide a statistical test for the key comparisons.
minor comments (5)
  1. [Table 4] The notation "IPG (w/o T)" is confusing; the table caption explains "w/ T" as weight transfer between tasks, but the main text in Sec. 5.3 discusses "when IPG is not used for weight transfer". Please make the table rows and the text consistent.
  2. [Table 15] The FLOPs entry for iDPA reads "506.00/501.00 GMac"; the slash is unexplained and should be resolved to a single value or clearly defined.
  3. [Sec. 4.2, Eq. (8)] The notation using \dot{p}_i and \ddot{p}_i is difficult to distinguish in print; please introduce clearer symbols for the intermediate activated prompt and the final prompt.
  4. [Fig. 2] Several components in Fig. 2 (e.g., the projection layers, the shared weights, and the exact flow from the prompt pool to the fusion encoder) are not described in enough detail in the text; a step-by-step caption or a short algorithm box would improve reproducibility.
  5. [Throughout] Dataset names contain typos (e.g., "Meneng" for Meningioma); please standardize names in tables and text.

Circularity Check

2 steps flagged · score 6.0 of 10

The DPA equivalence/convergence proof is circular: Lemma A.2 assumes the optimal representation is precisely the B branch that DPA keeps, so the conclusion that DPA matches full prompt attention is assumed rather than derived; the empirical SOTA claim is independently benchmarked but Appendix B admits reproduced prompt baselines may be handicapped.

  1. self definitional [Appendix A.3, Lemma A.2 (Eqs. 18-20)]
    "Lemma A.2. Let f1 and f2 be our models. Assume both models have converged to any local minima and there be an optimal representation f ∗ = B. ... Under convergence, M1A = 0, thus h(Concat[A; B]) = M2B = h′(B). ... y1 = h(Concat[A; B]) = h′(B) = y2. Thus f2 performs as well as f1 after convergence to local minima."

    The lemma is meant to prove that the decoupled branch retained by DPA (B, cf. Eq. 15: f2 = B) performs as well as the full two-branch attention output f1 = Concat[A; B]. That conclusion is built into the premises: f* is defined to be B, and convergence is used to assert M1A = 0, i.e., the deleted branch A has no effect. The proof therefore assumes the absence of the very term whose removal it claims to justify; the equality of f2 and f1 follows by construction, not by analysis.

  2. other [Section 4.3, Eqs. (10)-(13)]
    "By examining Eq. (10), we introduce the DPA mechanism, which decouples the knowledge transfer processes pv → ft and fv → ft in PA. This is achieved through: 1) discarding the learning of [ ¯pt] similar to VPT-Deep ... and 2) re-normalizing the attention weights through the weight adjustment scalar λ(ft) ... [ ¯ft = Attnv→t(fv, ft) + λ(ft)/(1−λ(ft)) Attnv→t(pv, ft) = Attnv→t(fv, ft) + λAttnv→t(pv, ft), where λ is a learnable parameter]."

    The move from Eq. (10) to Eq. (13) is presented as a derivation, but it is a design selection: the A block is discarded, the (1−λ(ft)) factor on the original branch is removed, and the attention-derived coefficient λ(ft)/(1−λ(ft)) is replaced by an independently learned scalar λ. None of these replacements follows from Eq. (10), so DPA is defined, not derived. The later Lemma A.2 closes the loop by assuming the discarded A branch is zero at convergence, exactly the point that would need to be established.

full rationale

The empirical claims of the paper rest on an external benchmark (ODinM-13) and are not fitted inputs renamed as predictions, so the SOTA comparison itself is not circular. The circularity lies in the theoretical support for DPA: the appendix's convergence proof (Lemma A.2) assumes an optimal representation f* = B, which is precisely the branch DPA keeps, and additionally assumes M1A = 0 so the discarded branch vanishes; the conclusion that f2 matches f1 is therefore contained in the assumptions. Similarly, Eqs. (10)-(13) in Section 4.3 do not derive DPA from prompt attention; they define it by dropping branch A and substituting a learnable λ, and the proof then assumes that dropped branch was irrelevant. No load-bearing self-citation chain or imported uniqueness theorem is present; self-citations (e.g., Yi et al. 2023, Guo et al. 2023) are contextual only. Separately, the paper's own Appendix B admits that prompt-based baselines were reproduced by prompting all layers rather than their original configurations, which may disadvantage them; that admission undermines the fairness of the numerical comparison but is a correctness/fairness issue rather than circularity. Overall: partial circularity in the derivation chain, with an independent empirical benchmark, so a moderate score of 6 is appropriate.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The empirical claim rests on a frozen GLIP backbone, a 13-dataset benchmark, and several hyperparameters tuned on that benchmark. The theoretical appendix adds an unjustified, circular optimality assumption. No new physical or ontological entities are introduced; IPG and DPA are architectural modules.

free parameters (7)
  • γ (RoI expansion factor) = 1.30 (Table 7; Eq. 6 text says 1.32)
    Selected by grid search on ODinM-13 to maximize FAP (Table 7).
  • Prompt length l = 10
    Chosen from sweep in Fig. 5d as best trade-off across full and few-shot settings.
  • Learning rate = 0.01 (Appendix B) or 0.1 (Sec. 5.1 text)
    Grid search reported in Fig. 5c; main text and appendix disagree.
  • Number of instances per class M = 1000 (full data), m (few-shot)
    Set to ensure diverse concept coverage; not tuned.
  • λ (DPA scale) = learnable vector in R^{1xd}, initialized 0
    Introduced as a learnable replacement for the attention weight in Eq. (13); fitted during training.
  • α, τ (CCPKI scale and temperature) = learnable
    Learnable parameters in Eq. (8) modulating concept activation.
  • Number of X-Attn layers = 6 (all)
    Selected from Table 6 as best performance; fewer layers trade off accuracy for speed.
assumptions (6)
  • standard math Block softmax decomposition of attention over concatenated prompt and token sequences (Eq. 10)
    Standard property of softmax attention used to expand Attn([pv;fv],[pt;ft]).
  • domain assumption GLIP pre-trained on natural-domain data is a sufficient frozen base for medical detection
    The entire method freezes GLIP and only tunes prompts; if GLIP features cannot encode medical concepts, the approach fails. Stated as a starting point in Sec. 4.1 and 5.1.
  • domain assumption ODinM-13 task order and few-shot splits are representative and fair
    Results are averaged over 3 seeds with task order tied to seed, but the choice of datasets and order is not justified beyond coverage of modalities and organs. Sec. 5.1 and Appendix B.
  • ad hoc to paper The optimal representation after convergence is exactly branch B (f* = B)
    Lemma A.2 in Appendix A assumes this; it is the conclusion of the proof, making the proof circular.
  • ad hoc to paper The head h is locally linear around the optimum and M1A = 0 at convergence
    Unproven assumption in Appendix A.3 used to conclude y1 = y2; not generally true for neural network heads.
  • ad hoc to paper Reproducing baseline prompts in all encoder layers is a fair comparison
    Original L2P/DualPrompt etc. use different prompt placements; Appendix B acknowledges the discrepancy may cause suboptimal baseline performance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of iDPA: Instance Decoupled Prompt Attention for Incremental Medical Object Detection." pith.science (2026). https://pith.science/paper/LE4MAGRM

@misc{pith2026250600406,
  author       = {Pith},
  title        = {Pith review of: iDPA: Instance Decoupled Prompt Attention for Incremental Medical Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LE4MAGRM}},
  note         = {Machine review of arXiv:2506.00406}
}
read the original abstract

Existing prompt-based approaches have demonstrated impressive performance in continual learning, leveraging pre-trained large-scale models for classification tasks; however, the tight coupling between foreground-background information and the coupled attention between prompts and image-text tokens present significant challenges in incremental medical object detection tasks, due to the conceptual gap between medical and natural domains. To overcome these challenges, we introduce the \method~framework, which comprises two main components: 1) Instance-level Prompt Generation (\ipg), which decouples fine-grained instance-level knowledge from images and generates prompts that focus on dense predictions, and 2) Decoupled Prompt Attention (\dpa), which decouples the original prompt attention, enabling a more direct and efficient transfer of prompt information while reducing memory usage and mitigating catastrophic forgetting. We collect 13 clinical, cross-modal, multi-organ, and multi-category datasets, referred to as \dataset, and experiments demonstrate that \method~outperforms existing SOTA methods, with FAP improvements of 5.44\%, 4.83\%, 12.88\%, and 4.59\% in full data, 1-shot, 10-shot, and 50-shot settings, respectively.

Figures

Figures reproduced from arXiv: 2506.00406 by the authors.

Figure 1
Figure 1. Comparison of our method (iDPA) with prior methods, highlighting improved object localization and recognition through instance-level prompt generation and decoupled prompt attention for medical detection tasks. Vision Language Object Detection (VLOD) (Li et al., 2022; Liu et al., 2023a; Cheng et al., 2024), a paradigm that enables recognition of novel categories and scenes, has advanced object detection (OD) by inte… view at source ↗
Figure 2
Figure 2. Overview of iDPA. Based on a frozen pre-trained VLOD model with visual-language interaction modules (e.g., GLIP (Li et al., 2022)), iDPA integrates Instance-level Prompt Generation (IPG) and Decoupled Prompt Attention (DPA) to enhance object localization and recognition, optimizing knowledge transfer for medical detection tasks. 4. Methodology 4.1. Overview of iDPA To effectively achieve the IMOD goal, our core idea… view at source ↗
Figure 3
Figure 3. Comparison between Decoupled Prompt Attention [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Performance variation of different CL methods [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Impact of context knowledge, scale types, learning rate, and prompt length on model performance. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Performance variation of iDPA in different location. [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Visualization results of iDPA compared with L2P and Zero-shot at the end of training with random seed 0 on [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 36 canonical work pages

  1. [1]

    Flamingo: a visual language model for few-shot learning

    Alayrac, J.-B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35: 0 23716--23736, 2022

    2022

  2. [2]

    Memory aware synapses: Learning what (not) to forget

    Aljundi, R., Babiloni, F., Elhoseiny, M., Rohrbach, M., and Tuytelaars, T. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European conference on computer vision (ECCV), pp.\ 139--154, 2018

    work page 2018

  3. [3]

    J., Fern \'a ndez-Esparrach, G., Gil, D., Rodr \' guez, C., and Vilari \ n o, F

    Bernal, J., S \'a nchez, F. J., Fern \'a ndez-Esparrach, G., Gil, D., Rodr \' guez, C., and Vilari \ n o, F. Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Computerized medical imaging and graphics, 43: 0 99--111, 2015

    work page 2015

  4. [4]

    C., Collins, D

    Boccardi, M., Bocchetta, M., Morency, F. C., Collins, D. L., Nishikawa, M., Ganzola, R., Grothe, M. J., Wolf, D., Redolfi, A., Pievani, M., et al. Training labels for hippocampal segmentation based on the eadc-adni harmonized hippocampal protocol. Alzheimer's & Dementia, 11 0 (2): 0 175--183, 2015

    work page 2015

  5. [5]

    End-to-end object detection with transformers

    Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., and Zagoruyko, S. End-to-end object detection with transformers. In European conference on computer vision, pp.\ 213--229. Springer, 2020

    2020

  6. [6]

    D., Pappachan, J

    Cassidy, B., Reeves, N. D., Pappachan, J. M., Gillespie, D., O’Shea, C., Rajbhandari, S., Maiya, A. G., Frank, E., Boulton, A. J., Armstrong, D. G., et al. The dfuc 2020 dataset: Analysis towards diabetic foot ulcer detection. touchREVIEWS in Endocrinology, 17 0 (1): 0 5, 2021

    work page 2020

  7. [7]

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

    Cheng, T., Song, L., Ge, Y., Liu, W., Wang, X., and Shan, Y. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 16901--16911, 2024

    work page 2024

  8. [8]

    Zero-shot generalizable incremental learning for vision-language object detection

    Deng, J., Zhang, H., Ding, K., Hu, J., Zhang, X., and Wang, Y. Zero-shot generalizable incremental learning for vision-language object detection. arXiv preprint arXiv:2403.01680, 2024

    arXiv 2024

Show all 61 references
  1. [9]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Devlin, J. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018

    2018 arXiv

  2. [10]

    Don't stop learning: Towards continual learning for the clip model

    Ding, Y., Liu, L., Tian, C., Yang, J., and Ding, H. Don't stop learning: Towards continual learning for the clip model. arXiv preprint arXiv:2207.09248, 2022

    2022 arXiv

  3. [11]

    Dytox: Transformers for continual learning with dynamic token expansion

    Douillard, A., Ram \'e , A., Couairon, G., and Cord, M. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 9285--9295, 2022

    2022

  4. [12]

    Multi-task learning for thyroid nodule segmentation with thyroid region prior

    Gong, H., Chen, G., Wang, R., Xie, X., Mao, M., Yu, Y., Chen, F., and Li, G. Multi-task learning for thyroid nodule segmentation with thyroid region prior. In 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI), pp.\ 257--261. IEEE, 2021

    2021

  5. [13]

    Multiple prompt fusion for zero-shot lesion detection using vision-language models

    Guo, M., Yi, H., Qin, Z., Wang, H., Men, A., and Lao, Q. Multiple prompt fusion for zero-shot lesion detection using vision-language models. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pp.\ 283--292. Springer, 2023

    2023

  6. [14]

    Towards a unified view of parameter-efficient transfer learning

    He, J., Zhou, C., Ma, X., Berg-Kirkpatrick, T., and Neubig, G. Towards a unified view of parameter-efficient transfer learning. arXiv preprint arXiv:2110.04366, 2021

    2021 arXiv

  7. [15]

    H., Riegler, M

    Jha, D., Smedsrud, P. H., Riegler, M. A., Halvorsen, P., Lange, T. d., Johansen, D., and Johansen, H. D. Kvasir-seg: A segmented polyp dataset. In International Conference on Multimedia Modeling, pp.\ 451--462. Springer, 2020

    2020

  8. [16]

    Video polyp segmentation: A deep learning perspective

    Ji, G.-P., Xiao, G., Chou, Y.-C., Fan, D.-P., Zhao, K., Chen, G., and Van Gool, L. Video polyp segmentation: A deep learning perspective. Machine Intelligence Research, 19 0 (6): 0 531--549, 2022

    2022

  9. [17]

    Visual prompt tuning

    Jia, M., Tang, L., Chen, B.-C., Cardie, C., Belongie, S., Hariharan, B., and Lim, S.-N. Visual prompt tuning. In European Conference on Computer Vision, pp.\ 709--727. Springer, 2022

    2022

  10. [18]

    Kim, B., Yu, J., and Hwang, S. J. Eclipse: Efficient continual learning in panoptic segmentation with visual prompt tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 3346--3356, 2024

    2024

  11. [19]

    A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al

    Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A. A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114 0 (13): 0 3521--3526, 2017

    2017

  12. [20]

    A two-stream continual learning system with variational domain-agnostic feature replay

    Lao, Q., Jiang, X., Havaei, M., and Bengio, Y. A two-stream continual learning system with variational domain-agnostic feature replay. IEEE Transactions on Neural Networks and Learning Systems, 33 0 (9): 0 4466--4478, 2021 a

    2021

  13. [21]

    Focl: Feature-oriented continual learning for generative models

    Lao, Q., Mortazavi, M., Tahaei, M., Dutil, F., Fevens, T., and Havaei, M. Focl: Feature-oriented continual learning for generative models. Pattern Recognition, 120: 0 108127, 2021 b

    2021

  14. [22]

    Mixture of experts meets prompt-based continual learning

    Le, M., Nguyen, A., Nguyen, H., Nguyen, T., Pham, T., Van Ngo, L., and Ho, N. Mixture of experts meets prompt-based continual learning. arXiv preprint arXiv:2405.14124, 2024

    2024 arXiv

  15. [23]

    Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., and Hoi, S. C. H. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34: 0 9694--9705, 2021

    2021

  16. [24]

    H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.-N., et al

    Li, L. H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.-N., et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 10965--10975, 2022

    2022

  17. [25]

    Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting

    Li, X., Zhou, Y., Wu, T., Socher, R., and Xiong, C. Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In International conference on machine learning, pp.\ 3925--3934. PMLR, 2019

    2019

  18. [26]

    and Hoiem, D

    Li, Z. and Hoiem, D. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40 0 (12): 0 2935--2947, 2017

    2017

  19. [27]

    Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Doll \'a r, P., and Zitnick, C. L. Microsoft coco: Common objects in context. In Computer Vision--ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pp....

    2014

  20. [28]

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

    Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023 a

    2023 arXiv

  21. [29]

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

    Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Jiang, Q., Li, C., Yang, J., Su, H., et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pp.\ 38--55. Springer, 2025

    2025

  22. [30]

    Rethinking computer-aided tuberculosis diagnosis

    Liu, Y., Wu, Y.-H., Ban, Y., Wang, H., and Cheng, M.-M. Rethinking computer-aided tuberculosis diagnosis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 2646--2655, 2020

    2020

  23. [31]

    Continual detection transformer for incremental object detection

    Liu, Y., Schiele, B., Vedaldi, A., and Rupprecht, C. Continual detection transformer for incremental object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 23799--23808, 2023 b

    2023

  24. [32]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., and Guo, B. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 10012--10022, 2021

    2021

  25. [33]

    and Ranzato, M

    Lopez-Paz, D. and Ranzato, M. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30, 2017

    2017

  26. [34]

    Decoupled weight decay regularization

    Loshchilov, I. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  27. [35]

    Segment anything in medical images

    Ma, J., He, Y., Li, F., Han, L., You, C., and Wang, B. Segment anything in medical images. Nature Communications, 15 0 (1): 0 654, 2024

    2024

  28. [36]

    and Lazebnik, S

    Mallya, A. and Lazebnik, S. Packnet: Adding multiple tasks to a single network by iterative pruning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pp.\ 7765--7773, 2018

    2018

  29. [37]

    S., Hang, D

    Ngoc Lan, P., An, N. S., Hang, D. V., Long, D. V., Trung, T. Q., Thuy, N. T., and Sang, D. V. Neounet: Towards accurate colon polyp segmentation and neoplasm detection. In Advances in visual computing: 16th international symposium, ISVC 2021, virtual event, October 4-6, 2021, ...

    2021

  30. [38]

    Medical image understanding with pretrained vision language models: A comprehensive study

    Qin, Z., Yi, H., Lao, Q., and Li, K. Medical image understanding with pretrained vision language models: A comprehensive study. In The Eleventh International Conference on Learning Representations

  31. [39]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pp.\ 8748--8763. PMLR, 2021

    2021

  32. [40]

    Sam 2: Segment anything in images and videos

    Ravi, N., Gabeur, V., Hu, Y.-T., Hu, R., Ryali, C., Ma, T., Khedr, H., R \"a dle, R., Rolland, C., Gustafson, L., et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024

    2024 arXiv

  33. [41]

    Rebuffi, S.-A., Kolesnikov, A., Sperl, G., and Lampert, C. H. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pp.\ 2001--2010, 2017

    2001

  34. [42]

    Faster r-cnn: Towards real-time object detection with region proposal networks

    Ren, S. Faster r-cnn: Towards real-time object detection with region proposal networks. arXiv preprint arXiv:1506.01497, 2015

    2015 arXiv

  35. [43]

    Experience replay for continual learning

    Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T., and Wayne, G. Experience replay for continual learning. Advances in neural information processing systems, 32, 2019

    2019

  36. [44]

    Setio, A. A. A., Traverso, A., De Bel, T., Berens, M. S., Van Den Bogaard, C., Cerello, P., Chen, H., Dou, Q., Fantacci, M. E., Geurts, B., et al. Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: ...

    2017

  37. [45]

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

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

    2019

  38. [46]

    K., Kim, J., and Kim, J

    Shin, H., Lee, J. K., Kim, J., and Kim, J. Continual learning with deep generative replay. Advances in neural information processing systems, 30, 2017

    2017

  39. [47]

    Incremental learning of object detectors without catastrophic forgetting

    Shmelkov, K., Schmid, C., and Alahari, K. Incremental learning of object detectors without catastrophic forgetting. In Proceedings of the IEEE international conference on computer vision, pp.\ 3400--3409, 2017

    2017

  40. [48]

    S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., and Kira, Z

    Smith, J. S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., and Kira, Z. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Visio...

    2023

  41. [49]

    Mind the interference: Retaining pre-trained knowledge in parameter efficient continual learning of vision-language models

    Tang, L., Tian, Z., Li, K., He, C., Zhou, H., Zhao, H., Li, X., and Jia, J. Mind the interference: Retaining pre-trained knowledge in parameter efficient continual learning of vision-language models. In European Conference on Computer Vision, pp.\ 346--365. Springer, 2025

    2025

  42. [50]

    D., Graham, S., Kurc, T., To, M

    Vu, Q. D., Graham, S., Kurc, T., To, M. N. N., Shaban, M., Qaiser, T., Koohbanani, N. A., Khurram, S. A., Kalpathy-Cramer, J., Zhao, T., et al. Methods for segmentation and classification of digital microscopy tissue images. Frontiers in bioengineering and biotechnology, pp.\ 53, 2019

    2019

  43. [51]

    S-prompts learning with pre-trained transformers: An occam’s razor for domain incremental learning

    Wang, Y., Huang, Z., and Hong, X. S-prompts learning with pre-trained transformers: An occam’s razor for domain incremental learning. Advances in Neural Information Processing Systems, 35: 0 5682--5695, 2022 a

    2022

  44. [52]

    Dualprompt: Complementary prompting for rehearsal-free continual learning

    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 European Conference on Computer Vision, pp.\ 631--648. Springer, 2022 b

    2022

  45. [53]

    Learning to prompt for continual learning

    Wang, Z., Zhang, Z., Lee, C.-Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., and Pfister, T. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 139--149, 2022 c

    2022

  46. [54]

    W., Li, M., Kornblith, S., Roelofs, R., Lopes, R

    Wortsman, M., Ilharco, G., Kim, J. W., Li, M., Kornblith, S., Roelofs, R., Lopes, R. G., Hajishirzi, H., Farhadi, A., Namkoong, H., et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 7959--7...

    2022

  47. [55]

    Wu, S., Zhang, W., Jin, S., Liu, W., and Loy, C. C. Aligning bag of regions for open-vocabulary object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 15254--15264, 2023

    2023

  48. [56]

    Multi-modal queried object detection in the wild

    Xu, Y., Zhang, M., Fu, C., Chen, P., Yang, X., Li, K., and Xu, C. Multi-modal queried object detection in the wild. Advances in Neural Information Processing Systems, 36, 2024

    2024

  49. [57]

    Continual self-supervised learning: Towards universal multi-modal medical data representation learning

    Ye, Y., Xie, Y., Zhang, J., Chen, Z., Wu, Q., and Xia, Y. Continual self-supervised learning: Towards universal multi-modal medical data representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 11114--11124, 2024

    2024

  50. [58]

    Towards general purpose medical ai: Continual learning medical foundation model

    Yi, H., Qin, Z., Lao, Q., Xu, W., Jiang, Z., Wang, D., Zhang, S., and Li, K. Towards general purpose medical ai: Continual learning medical foundation model. arXiv preprint arXiv:2303.06580, 2023

    2023 arXiv

  51. [59]

    Yoon, J., Yang, E., Lee, J., and Hwang, S. J. Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547, 2017

    2017 arXiv

  52. [60]

    Medical sam 2: Segment medical images as video via segment anything model 2

    Zhu, J., Hamdi, A., Qi, Y., Jin, Y., and Wu, J. Medical sam 2: Segment medical images as video via segment anything model 2. arXiv preprint arXiv:2408.00874, 2024

    2024 arXiv

  53. [61]

    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 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.