Demystifying Catastrophic Forgetting in Two-Stage Incremental Object Detector

De Cheng; Di Xu; Peng Wang; Qirui Wu; Shizhou Zhang; Yanning Zhang; Yinghui Xing

REVIEW 3 major objections 4 minor 19 references

Demystifying Catastrophic Forgetting in Two-Stage Incremental Object Detector

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Catastrophic forgetting in two-stage detection is localized to the RoI Head classifier.

desk verdict A useful dissection of where Faster R-CNN forgets, but the headline claim conflates classifier weight loss with feature drift; the method itself is solid and worth engaging. read the letter →

arxiv 2502.05540 v3 pith:ITLKNZFO submitted 2025-02-08 cs.CV

Qirui Wu , Shizhou Zhang , De Cheng , Yinghui Xing , Di Xu , Peng Wang , Yanning Zhang This is my paper · ORCID

classification cs.CV

keywords catastrophicforgettingincrementalobjectdetectionFasterR-CNNRoIheadclassifierprototypereplaynullspacegradientprojectioncontinuallearningpseudo-labeling

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's central claim is that catastrophic forgetting in a two-stage incremental object detector is not distributed across the network: the region proposal network keeps its recall, the box regression branch keeps its accuracy, and nearly all of the forgetting sits in the RoI Head classifier. The authors support this with a dissection of Faster R-CNN that fixes proposals and classifier outputs independently, isolating each branch's contribution to the mAP drop. They then argue that this localization justifies a simpler remedy: replay per-class prototypes into the classification branch while projecting feature-extractor gradient updates into the null space of previous inputs to keep features and prototypes aligned. If the claim is right, incremental detection methods should spend their anti-forgetting effort on the classifier head rather than on the proposal or regression modules.

What carries the argument

The diagnostic mechanism is the manipulation of a fixed set of proposals: evaluating model M_i on a test set with proposals from later models isolates RPN forgetting, and fixing the classifier output to that of an earlier model isolates the regression branch's forgetting from the classifier's. The remedy's machinery is RePRE, which stores one coarse prototype per old class plus K fine-grained prototypes chosen by a density-aware greedy hypersphere selection, and replays them through the classifier softmax with a cross-entropy replay loss; and NSGP, which accumulates the uncentered covariance of old inputs, retains the null-space singular vectors, and multiplies each gradient by the projection matrix $B = U'U'^\top$ so updates are orthogonal to old-task inputs. Together they target precisely the diagnosed weak component.

What would settle it

Run the same fixed-proposals, fixed-classifier dissection on the COCO 40-40 or 70-10 splits: if the mAP drop of old classes is not dominated by the classifier branch (e.g., the regression branch alone causes a comparable drop when the classifier output is fixed to a frozen model), the claim that forgetting is predominantly localized to the RoI Head classifier is falsified. A second check is to measure RPN recall drop on COCO and the mAP impact of substituting later-stage proposals; if that impact clearly exceeds the 1.3% observed on VOC, the 'RPN forgets negligibly' leg fails.

Watch

Extended reading notes

Core claim

On the paper's own terms: sequential training of Faster R-CNN on Pascal VOC splits (5+5+5+5) shows three stable results: RPN recall curves are nearly unchanged across stages; swapping in later-stage proposals for the first-stage RPN costs only about 1.3% mAP; and fixing the classifier's predictions while letting the regressor vary across models yields an almost flat mAP, while letting the classifier vary causes a rapid drop. The conclusion is that catastrophic forgetting in two-stage incremental object detection is predominantly a failure of the RoI Head classifier, and that the regressor's robustness explains why classification-style continual-learning findings carry over to detection. Based on this, the proposed NSGP-RePRE replays coarse regional prototypes (class-wise means of RoI features) and fine-grained prototypes (averages of density-selected hyperspheres) into the classifier, and NSGP projects backbone and neck gradient updates into the null space of old inputs so the stored prototypes remain aligned with current RoI features.

Load-bearing premise

The component-level diagnosis is measured on a single protocol — 4-stage PASCAL VOC 5+5+5+5 with ResNet-50 Faster R-CNN, pseudo-labeling, and the top 1,000 proposals — and the paper assumes that the same localization (RPN stable, regressor robust, classifier fragile) holds in every other reported setting.

Editorial extensions

If this is right

Incremental object detection methods can stop treating the whole detector as a single fragile object and instead allocate anti-forgetting resources to the RoI Head classifier, leaving the RPN and box regression branch with little or no special protection.
A small per-class budget of stored prototypes (ten per class in this implementation) is enough to preserve old-class classification knowledge, so instance buffers can be replaced by compact feature summaries.
Prototype replay combined with feature-drift control should remain competitive or superior across single-step and multi-step VOC and COCO protocols, with the largest gains on the base classes learned in the first stage.
The localization claim implies that whole-feature knowledge distillation and whole-logit regularization are over-constrained for two-stage detectors; constraining the classifier head plus a null-space projection should capture most of the available benefit.

Reading between the lines

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

The paper leaves open whether the same localization holds in one-stage detectors; if it does, replaying classification-head features could serve as a cheap anti-forgetting fix there too.
The diagnostic protocol itself (fixed proposals with fixed classifier outputs) could be adopted as a routine check for other continual-learning vision models, revealing which component actually forgets before a mitigation is chosen.
The authors attribute the regressor's stability to an absence of task conflict in regression; that explanation implies a testable boundary, namely that a detector with class-conditioned box regression might display regression forgetting that the current two-branch design hides.
Because the anatomy experiments use pseudo-labeling, the diagnosis is conditional on that strategy; without pseudo-labeling, the foreground-shift problem might stress the regressor more than the reported experiments show.

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.

Desk Editor's Note

read the letter

Two things to know. First, the component-level analysis is genuinely new and mostly well-executed: RPN recall is stable, and the regressor is robust while the classification side degrades. Second, the paper's central attribution—'forgetting is predominantly localized to the RoI Head classifier'—is not actually isolated from feature-extractor drift. In Figure 3, fixed proposals are fed through the RoIAlign of later models, so the classifier sees drifted features. The drop attributed to classifier weights could be caused by backbone/neck drift. The dashed-line control using M1's labels shows the regressor is less sensitive, not that classifier weights are the locus.

The paper's own ablation makes this concrete. In Table 4, NSGP alone (feature-drift control) lifts old-class mAP from 46.6 to 62.3, while coarse prototype replay alone reaches only 49.8. The most effective intervention targets feature drift, not classifier replay. That does not kill the paper—the method combines both and works—but it substantially weakens the mechanistic claim. The authors even say the reason is unclear, which is honest.

What is good: the RPN stability and regressor robustness results are clean and useful. The proposed NSGP-RePRE is simple, reproducible (code released), and achieves strong old-class retention on VOC and COCO, often beating prior work on base classes. The ablations are informative. The soft spots: no error bars or multiple runs, so gains could be within noise; the 'state-of-the-art' claim is too broad given new-class results sometimes trail ABR/BPF; the anatomy is only measured on one architecture and a 4-stage VOC split, with generalization assumed.

The citation pattern is fine: NSGP builds on prior null-space methods, including the authors' own, and that is legitimate. The main issue is interpretational, not technical. A serious revision should add a controlled experiment that holds features fixed (e.g., using M1's backbone to extract RoI features for all later classifiers) to separate feature drift from classifier weight decay.

Who this is for: anyone working on incremental object detection. The method is a strong baseline, and the anatomy, even with the confound, redirects attention to the RoI head and feature drift. I'd take the result seriously but not at face value. It deserves peer review, with the confound and missing error bars as major revision points.

Referee Report

3 major / 4 minor

Summary. The paper analyzes catastrophic forgetting in two-stage incremental object detectors, focusing on Faster R-CNN. Through controlled experiments on PASCAL VOC (Figures 1–3), the authors conclude that forgetting is predominantly localized to the RoI Head classifier, while the RPN and the box regressor retain knowledge across incremental stages. Based on this finding, they propose NSGP-RePRE, which combines Regional Prototype Replay (coarse and fine-grained prototypes) for the RoI classifier with Null Space Gradient Projection to stabilize the feature extractor. The method is evaluated on PASCAL VOC and MS COCO under multiple incremental settings, reporting improvements over prior approaches. The paper is well structured and presents a useful empirical study, but the central localization claim is not fully isolated from feature-extractor drift, and the method's own ablation results complicate the narrative.

Significance. If the central claim is correct, the paper provides a clear and actionable insight: incremental object detection methods should direct anti-forgetting resources to the RoI classification head rather than to the RPN or the regression branch. The proposed NSGP-RePRE is simple and achieves strong results on standard benchmarks, and the authors make their code available. However, the evidence for the localization claim is currently inconclusive because the experimental design does not separate classifier-weight forgetting from feature drift, and the ablation study (Table 4) shows that null-space gradient projection alone yields much larger gains than classifier prototype replay, which is in tension with the claim that the classifier is the predominant forgetting source. The paper would be more convincing if the authors either strengthened the anatomy experiment with a feature-drift control or revised the claim to reflect that both classifier and feature drift contribute.

major comments (3)

[Section 3.3, Figure 3] The fixed-proposal experiment does not isolate RoI-head classifier weight forgetting from feature-extractor drift. When proposals Pi are fed into a later model Mj, RoIAlign extracts features using Mj's backbone/neck, so the classifier receives features that have drifted as a result of incremental updates. The drop in the solid line could therefore be caused by a shift in the feature distribution rather than by changes to the classifier weights. The dashed-line control, which uses M1's classification outputs together with each model's regressor, only shows that the regressor is insensitive to feature drift; it does not control for the drift experienced by the classifier. To support the stated conclusion, the authors should evaluate the classifier of Mj on features extracted by Mi (for instance, by using a fixed feature extractor or by explicitly measuring and compensating for feature drift). Without such a control, the claim that forgetting is 'predominantly localized to the RoI Head classifier' is confounded.
[Table 4, ablation study] The ablation results undermine the paper's central claim. Variant (b), which adds NSGP alone to the baseline, improves old-class mAP (1-5) from 46.6 to 62.3, while variant (c), which adds coarse prototype replay alone, improves it only to 49.8. If the RoI classifier were the dominant source of forgetting, replaying prototypes to the classifier should produce a much larger effect than feature-drift regularization. The observed pattern suggests that feature drift is at least as important, if not more, which is inconsistent with the statement in the abstract that catastrophic forgetting is 'predominantly localized to the RoI Head classifier.' The authors need to either reconcile this discrepancy or substantially revise the claim to acknowledge that feature drift is a major component of forgetting.
[Tables 1–3] All experiments are reported from a single run without error bars or standard deviations. Given that several improvements over prior work are small (e.g., 0.3 AP on COCO 70-10 in Table 3, and 0.6–1.4 mAP differences in Table 1), it is unclear whether these differences are statistically significant. The authors should repeat the experiments with at least three random seeds and report the mean and variance for the main results. This is particularly important because the paper's contribution hinges on the reliability of the reported gains.

minor comments (4)

[Abstract and author affiliations] There are several typos, including 'chanllenge' in the abstract, 'Unviersity' in the author affiliation, and 'deacy' in Appendix A. These should be corrected.
[Section 5.2, Tables 1–3] The paper claims 'state-of-the-art performance' but on new-class partitions the method is sometimes worse than prior work (e.g., Table 1, 15-5: NSGP-RePRE 61.8 on new classes vs. BPF 63.0; Table 1, 5-15: 74.5 vs. 75.3). The authors should more precisely state that the gains are primarily on base classes and discuss the trade-off in plasticity.
[Appendix D, Table 5] The discussion of B versus B' normalization is a useful design choice, but the motivation for applying B' only to the backbone is not fully justified. The authors note that lower update stride leads to slower optimization, but the table shows a large drop for RoI Head when B' is used (41.9 vs. 63.2). A brief explanation of why the RoI Head is particularly sensitive would help the reader understand this asymmetric choice.
[Section 4.1, Eq. (7)] The replay loss notation is slightly ambiguous: the summation over fine-grained prototypes reuses the same class label yk for each prototype. It would be clearer to write y_{k} for the coarse prototype and y_{k,j} for the fine-grained prototypes, even though they share the same class index.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; the central forgetting-localization claim is an empirical measurement, with only a minor non-load-bearing self-citation for NSGP's nullity selection.

full rationale

The central claim—that catastrophic forgetting in Faster R-CNN is predominantly localized to the RoI Head classifier while the regressor retains robustness—is established by the anatomy experiments in Section 3, not by a derivation from an assumed conclusion. Section 3.3 fixes proposals from M_i and compares RoI Head outputs of later models; while the RoIAlign features are extracted from later models' backbone/neck, so feature drift is a genuine confound for the locus of forgetting, this is an interpretation/correctness risk rather than circularity. No equation defines 'classifier forgetting' as the measured drop, and the conclusion is not baked into the protocol. The NSGP component borrows null-space gradient projection from Wang et al. (2021) and the adaptive nullity-selection rule from Lu et al., a prior work by the same research group; this is a minor self-citation supplying a hyperparameter-selection scheme, but it does not define or force the paper's target result, and the method is evaluated against external benchmarks and prior methods. The central anatomy finding stands independently of the self-cited module. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported, and the paper explicitly notes that the reason for the classifier/regressor asymmetry is unclear (Section 3.3), further confirming that this is an empirical observation rather than a self-referential derivation. Overall, the derivation chain is self-contained with respect to its central claim; the score of 2 reflects only the non-load-bearing self-citation in Appendix D.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or architectural entities; its RePRE prototypes are data structures, and NSGP is an adaptation of published null-space projection methods. The free parameters are mainly the prototype generation hyperparameters and the imported NSGP nullity selection. The main assumptions are the representativeness of the VOC 5+5+5+5 anatomy and the sufficiency of accumulated uncentered covariance for null-space construction.

free parameters (4)

Number of fine-grained prototypes per class (K) = 9 (10 prototypes per class total)
Set to 9 in implementation (Appendix A); no ablation over K is reported, so the value is chosen by hand and affects replay loss magnitude and storage.
Hypersphere radius r for cosine-similarity neighborhood = 0.6
Set to 0.6 in Appendix A; controls which RoI features are aggregated into fine-grained prototypes and is not swept.
NSGP null-space dimension R = Adaptive, following (Lu et al.)
The number of retained singular vectors is selected by an adaptive strategy imported from the authors' prior continual-learning work; the criterion is not fully described in this paper, so the effective null-space size is an external input.
Pseudo-labeling threshold or selection rule = As in BPF (Mo et al., 2024)
The pseudo-labeling implementation is inherited from BPF but its confidence threshold or selection rule is not specified, and it affects the baseline and all variants.

assumptions (4)

domain assumption Component-wise forgetting in Faster R-CNN is stable across class orderings, datasets, and step counts; the VOC 5+5+5+5 anatomy is representative.
The paper's key insight is derived from one architecture (R50 Faster R-CNN), one dataset split, and one proposal count (Sections 3.2-3.3), then used to justify method design and SOTA claims across many settings (Section 5.2).
domain assumption Uncentered covariance of layer inputs (Eq. 11-12) is a sufficient statistic for the null space of old inputs.
NSGP builds projection matrix B from accumulated uncentered covariance and SVD (Eqs. 13-15); if this approximation misses directions, RoI features can drift and prototypes misalign.
ad hoc to paper The adaptive nullity selection of (Lu et al.) and the B versus B' normalization choice transfer to Faster R-CNN.
Appendix D imports the singular-value selection strategy from the authors' own prior work and justifies B versus B' empirically (Table 5); neither choice is derived from first principles.
domain assumption Pseudo-labeling resolves the missing-annotation problem well enough that component-level forgetting measurements are not dominated by label noise.
The anatomy and all variants use pseudo-labeling as in BPF (Section 5.1), so the measured forgetting includes any bias introduced by imperfect pseudo-labels.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Demystifying Catastrophic Forgetting in Two-Stage Incremental Object Detector." pith.science (2026). https://pith.science/paper/ITLKNZFO

@misc{pith2026250205540,
  author       = {Pith},
  title        = {Pith review of: Demystifying Catastrophic Forgetting in Two-Stage Incremental Object Detector},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ITLKNZFO}},
  note         = {Machine review of arXiv:2502.05540}
}

read the original abstract

Catastrophic forgetting is a critical chanllenge for incremental object detection (IOD). Most existing methods treat the detector monolithically, relying on instance replay or knowledge distillation without analyzing component-specific forgetting. Through dissection of Faster R-CNN, we reveal a key insight: Catastrophic forgetting is predominantly localized to the RoI Head classifier, while regressors retain robustness across incremental stages. This finding challenges conventional assumptions, motivating us to develop a framework termed NSGP-RePRE. Regional Prototype Replay (RePRE) mitigates classifier forgetting via replay of two types of prototypes: coarse prototypes represent class-wise semantic centers of RoI features, while fine-grained prototypes model intra-class variations. Null Space Gradient Projection (NSGP) is further introduced to eliminate prototype-feature misalignment by updating the feature extractor in directions orthogonal to subspace of old inputs via gradient projection, aligning RePRE with incremental learning dynamics. Our simple yet effective design allows NSGP-RePRE to achieve state-of-the-art performance on the Pascal VOC and MS COCO datasets under various settings. Our work not only advances IOD methodology but also provide pivotal insights for catastrophic forgetting mitigation in IOD. Code is available at \href{https://github.com/fanrena/NSGP-RePRE}{https://github.com/fanrena/NSGP-RePRE} .

Figures

Figures reproduced from arXiv: 2502.05540 by the authors.

**Figure 2.** Results of Mi on Di with different proposals. Pj are produced by corresponding Mj . 1 2 3 4 Model 40 60 80 mAP (%) Eval with fixed P1 (a) On Dtest 1 2 3 4 Model 70 80 mAP (%) Eval with fixed P2 (b) On Dtest 2 3 4 Model 75 80 mAP (%) Eval with fixed P3 (c) On Dtest 3 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Results of Mi on various Di by using a fixed set of proposals. “- -” indicates the classification results of each proposal is designated by Model freshly trained on the corresponding D. “—” indicates the predicted classification results for the corresponding model in the x-axis. 3.3. Anatomy of Faster R-CNN RPN’s recall ability remains consistent across sequential tasks. RPN allows the detector to generate possible… view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: The overall architecture of our NSGP-RePRE framework. This framework incorporates RePRE to mitigate [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: mAP of different model on Dtest 1 in VOC(5-5) settings. To better demonstrate the impact of our method on the classifier, P1 is fixed to all models. Fixed cls indicates the models classification results is designated by M1. achieves the highest performance among all mo…

**Figure 6.** Figure 6: Plot: Results of Mjoint after removing high-quality proposals with varying IoU threshold. Bar: The distribution of the proposals generated with Mjoint over IoU. The number on the bar indicates the count of proposals. A robust RoI head is capable of effectively offsetti…

**Figure 7.** Figure 7: FC/Conv FC/Conv Null Space Old Input Projection Forward Null Space [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 10 canonical work pages

[1]

L., and Kanan, C

Acharya, M., Hayes, T. L., and Kanan, C. Rodeo: Replay for online object detection. arXiv preprint arXiv:2008.06439,

arXiv 2008
[6]

Condensed Prototype Replay for Class Incremental Learning

Kong, J., Zong, Z., Zhou, T., and Shao, H. Condensed prototype replay for class incremental learning. arXiv preprint arXiv:2305.16143,

work page Pith review arXiv
[7]

Lin, T.-Y ., Maire, M., Belongie, S., Hays, J., Perona, P., Ra- manan, D., Doll´ar, 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. 740–

2014
[13]

Zhou, K., Yang, J., Loy, C

doi: 10.1109/TMM.2024.3521681. Zhou, K., Yang, J., Loy, C. C., and Liu, Z. Learning to prompt for vision-language models. International Jour- nal of Computer Vision, 130(9):2337–2348,

work page arXiv 2024
[14]

Lifelong object detection

Zhou, W., Chang, S., Sosa, N., Hamann, H., and Cox, D. Lifelong object detection. arXiv preprint arXiv:2009.01129,

arXiv 2009
[15]

On PASCAL VOC dataset, we train the network with SGD optimizer, momentum of 0.9 and weight decay of 10e-4

backbone pre-trained in ImageNet (Russakovsky et al., 2015). On PASCAL VOC dataset, we train the network with SGD optimizer, momentum of 0.9 and weight decay of 10e-4. We use a learning rate of 0.02 for all tasks. For MS COCO, we adopt AdamW as the optimizer, weight deacy of 0.01 and learning rate of 5e-5. Batch size is set to 16 for both datasets. In NSG...

work page 2015
[16]

To obtain the projection matrix of an FC layer or a convolution layer with parametersW , we first compute the uncentered covariance of Xt

FC/Conv FC/Conv Null Space Projection Old Input Forward Null Space Figure 7: An overview of NSGP. To obtain the projection matrix of an FC layer or a convolution layer with parametersW , we first compute the uncentered covariance of Xt. Specifically, we can accumulate uncentered covariance matrix in t-th stage Tt as: Tt = 1 Nt − 1 NtX i=1 x⊤ t,ixt,i. (11)...

work page 2021
[18]

68.7 67.1 67.9 64.7 56.4 58.4 67.0 58.1 62.6 68.7 56.7 65.7 62.0 55.7 58.9 FasterILOD (Ren et al., 2016)68.3 57.9 63.1 55.7 16.0 25.9 64.2 48.6 56.4 66.9 44.5 61.3 52.9 41.5 47.2 MMA (Cermelli et al., 2022)66.7 61.8 64.2 62.3 31.2 38.9 65.0 53.1 59.1 68.3 54.3 64.1 59.2 48.3 53.8 BPF (Mo et al.,

work page 2016

Show all 19 references

[2007]

The best performance in each is presented with bold, and the second best is presented with underline. 19-1 15-5 10-10 5-15 Method 1-19 20 1-20 Avg 1-15 16-20 1-20 Avg 1-10 11-20 1-20 Avg 1-5 5-15 1-20 Avg Joint 76.4 76.4 76.4 76.4 78.3 70.7 76.4 74.5 76.9 76.0 76.4 76.4 73.6 7...

2022
[2010]

A., Pritzel, A., and Wierstra, D

Fernando, C., Banarse, D., Blundell, C., Zwols, Y ., Ha, D., Rusu, A. A., Pritzel, A., and Wierstra, D. Pathnet: Evolu- tion channels gradient descent in super neural networks. arXiv preprint arXiv:1701.08734,

arXiv
[2015]

A., Rabinowitz, N

Rusu, A. A., Rabinowitz, N. C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., and Had- sell, R. Progressive neural networks. arXiv preprint arXiv:1606.04671,

arXiv
[2016]

Ovor: Oneprompt with virtual outlier regularization for rehearsal-free class- incremental learning

Huang, W.-C., Chen, C.-F., and Hsu, H. Ovor: Oneprompt with virtual outlier regularization for rehearsal-free class- incremental learning. arXiv preprint arXiv:2402.04129,

arXiv
[2017]

and Hussain, M

Khanam, R. and Hussain, M. Yolov11: An overview of the key architectural enhancements. arXiv preprint arXiv:2410.17725,

arXiv
[2020]

Augmented box replay: Overcoming foreground shift for incremental object detection

10 Demystifying Catastrophic Forgetting in Two-Stage Incremental Object Detector Yuyang, L., Yang, C., Dipam, G., Xialei, L., and van de Wei- jer, J. Augmented box replay: Overcoming foreground shift for incremental object detection. arXiv preprint arXiv:2307.12427,

arXiv
[2021]

Orthogonal subspace learning for language model continual learning

Wang, X., Chen, T., Ge, Q., Xia, H., Bao, R., Zheng, R., Zhang, Q., Gui, T., and Huang, X. Orthogonal subspace learning for language model continual learning. arXiv preprint arXiv:2310.14152, 2023a. Wang, Y ., Ma, Z., Huang, Z., Wang, Y ., Su, Z., and Hong, X. Isolation and im...

arXiv
[2022]

S., and Balasubramanian, V

Joseph, K., Khan, S., Khan, F. S., and Balasubramanian, V . N. Towards open world object detection. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 5830–5840, 2021a. Joseph, K., Rajasegaran, J., Khan, S., Khan, F. S., and Balasubraman...

arXiv
[2023]

Piggyback gan: Efficient lifelong learning for image conditioned generation

Zhai, M., Chen, L., He, J., Nawhal, M., Tung, F., and Mori, G. Piggyback gan: Efficient lifelong learning for image conditioned generation. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXI 16, pp. 397–413. Springer,

2020
[2024]

The best performance in each is presented with bold, and the second best is presented with underline

69.1 68.2 68.7 60.6 63.1 62.5 68.7 56.3 62.5 71.5 53.1 66.9 62.2 48.3 55.2 NSGP-RePRE-Coarse 71.9 66.2 69.1 65.9 63.8 64.3 68.7 54.8 61.8 77.0 53.9 71.2 71.2 50.6 60.9 NSGP-RePRE 72.4 67.6 70.0 64.6 66.1 65.7 70.1 58.8 64.4 77.7 55.0 72.0 69.9 55.1 62.5 Table 9: mAP results on...

2017
[2025]

doi: 10.1016/j.knosys

ISSN 0950-7051. doi: 10.1016/j.knosys. 2024.112920. URL https://doi.org/10.1016/ j.knosys.2024.112920. Yang, D., Zhou, Y ., Zhang, A., Sun, X., Wu, D., Wang, W., and Ye, Q. Multi-view correlation distillation for incremental object detection. Pattern Recognition, 131: 108863,

2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

L., and Kanan, C

Acharya, M., Hayes, T. L., and Kanan, C. Rodeo: Replay for online object detection. arXiv preprint arXiv:2008.06439,

arXiv 2008

[2] [6]

Condensed Prototype Replay for Class Incremental Learning

Kong, J., Zong, Z., Zhou, T., and Shao, H. Condensed prototype replay for class incremental learning. arXiv preprint arXiv:2305.16143,

work page Pith review arXiv

[3] [7]

Lin, T.-Y ., Maire, M., Belongie, S., Hays, J., Perona, P., Ra- manan, D., Doll´ar, 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. 740–

2014

[4] [13]

Zhou, K., Yang, J., Loy, C

doi: 10.1109/TMM.2024.3521681. Zhou, K., Yang, J., Loy, C. C., and Liu, Z. Learning to prompt for vision-language models. International Jour- nal of Computer Vision, 130(9):2337–2348,

work page arXiv 2024

[5] [14]

Lifelong object detection

Zhou, W., Chang, S., Sosa, N., Hamann, H., and Cox, D. Lifelong object detection. arXiv preprint arXiv:2009.01129,

arXiv 2009

[6] [15]

On PASCAL VOC dataset, we train the network with SGD optimizer, momentum of 0.9 and weight decay of 10e-4

backbone pre-trained in ImageNet (Russakovsky et al., 2015). On PASCAL VOC dataset, we train the network with SGD optimizer, momentum of 0.9 and weight decay of 10e-4. We use a learning rate of 0.02 for all tasks. For MS COCO, we adopt AdamW as the optimizer, weight deacy of 0.01 and learning rate of 5e-5. Batch size is set to 16 for both datasets. In NSG...

work page 2015

[7] [16]

To obtain the projection matrix of an FC layer or a convolution layer with parametersW , we first compute the uncentered covariance of Xt

FC/Conv FC/Conv Null Space Projection Old Input Forward Null Space Figure 7: An overview of NSGP. To obtain the projection matrix of an FC layer or a convolution layer with parametersW , we first compute the uncentered covariance of Xt. Specifically, we can accumulate uncentered covariance matrix in t-th stage Tt as: Tt = 1 Nt − 1 NtX i=1 x⊤ t,ixt,i. (11)...

work page 2021

[8] [18]

68.7 67.1 67.9 64.7 56.4 58.4 67.0 58.1 62.6 68.7 56.7 65.7 62.0 55.7 58.9 FasterILOD (Ren et al., 2016)68.3 57.9 63.1 55.7 16.0 25.9 64.2 48.6 56.4 66.9 44.5 61.3 52.9 41.5 47.2 MMA (Cermelli et al., 2022)66.7 61.8 64.2 62.3 31.2 38.9 65.0 53.1 59.1 68.3 54.3 64.1 59.2 48.3 53.8 BPF (Mo et al.,

work page 2016