IPSeg: Image Posterior Mitigates Semantic Drift in Class-Incremental Segmentation

Xiao Yu; Yan Fang; Yao Zhao; Yunchao Wei

REVIEW 4 major objections 6 minor 18 references

IPSeg: Image Posterior Mitigates Semantic Drift in Class-Incremental Segmentation

T0 review · 4 major / 6 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 IPSeg multiplies every pixel's class scores by image-level posterior probabilities learned alongside the segmentation heads, and reports that this rescaling plus a permanent/temporary semantics split curbs semantic drift — a 24.8-point…

desk verdict Solid, well-evidenced CISS paper with a genuinely new image-posterior mechanism; needs number corrections and a sharper per-class analysis of the image posterior branch. read the letter →

arxiv 2502.04870 v1 pith:KPGFLORI submitted 2025-02-07 cs.CV

Xiao Yu , Yan Fang , Yao Zhao , Yunchao Wei This is my paper · ORCID

classification cs.CV

keywords class-incrementalsemanticsegmentationdriftimageposteriorguidanceseparateoptimizationpseudo-labelnoisesemanticsdecouplingcatastrophicforgettingcontinuallearning

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

Class-incremental semantic segmentation is the setting where one model must keep segmenting already-learned classes while absorbing new ones from a stream of tasks. The paper's target is semantic drift: a pixel that meant 'background' in an earlier task can become a target object later, and models that freeze each task's prediction head produce heads with misaligned probability scales that mistake similar-looking classes, such as cow and horse. IPSeg claims that two mechanisms curb this drift: multiplying every pixel's class scores by an image-level posterior probability from a branch trained alongside the segmentation heads, and decoupling the learning targets into permanent semantics (background and unknown objects, learned continuously) and temporary semantics (each task's own classes, learned per phase). The reported evidence is consistent gains over prior state of the art on Pascal VOC 2012 and ADE20K, with the largest margins in long-horizon scenarios, including a 24.8-point improvement over SSUL-M on new classes in the VOC 2-2 setting. If the claim is right, a comparatively cheap global signal can keep pixel-level heads aligned across many incremental steps, which matters because long sequences of small tasks are where existing methods collapse.

What carries the argument

The load-bearing object is the product rule of Eq. 3: $p_i = \mathrm{Concat}(\alpha_{BC}, \sigma(\psi(h_\theta(x_i)))) \cdot \sigma(\phi_{0:T}(h_\theta(x_i)))$, which rescales every pixel's class scores by the image-level posterior. The image posterior branch $\psi$ is a pooling-plus-MLP image classifier appended to the shared backbone, trained incrementally with a BCE loss whose target is the ground-truth image label unioned with image-level pseudo-labels taken from the previous task heads (Eq. 2), and it is justified by the paper's Fact that an image's class domain is identical at image level and pixel level. Around that product the method wraps a label scheme (Eq. 5) that splits each training image into permanent semantics — pure background $c'_b$ and unknown foreground $c'_u$, assigned to a branch $\phi_p$ that trains across all phases — and temporary semantics — current target classes $C_t$ plus other foreground $c_f$, assigned to the per-task branch $\phi_t$ — together with a noise-filtering rule (Eq. 9, $\alpha_{NF}=0.4$) that suppresses $c_f$ votes when they outscore true target classes. The posterior branch supplies the global consistency check; the decoupling ensures that pseudo-label noise is not learned by the same parameters that must deliver clean predictions.

What would settle it

Compare, per class, the posterior branch's image-level recall against the pixel heads' accuracy on images where the class occupies a small pixel fraction or strongly resembles another class; wherever the posterior's recall is lower, the product rule of Eq. 3 suppresses correct pixels and the method would lose to its un-corrected baseline on those images. A sharper version replaces the learned posterior at inference with the ground-truth image labels (oracle) and with a deliberately corrupted posterior: the size of the oracle gain and the slope of the corruption response bracket how much of the reported 24.8-point improvement is a true rescaling effect versus inherited from the posterior branch's own accuracy.

Watch

Extended reading notes

Core claim

The paper argues that semantic drift in class-incremental segmentation has two causes that earlier work underplayed. The first, separate optimization, is inherent to architecture-based incremental models: each task head $\phi_t$ is trained only on its own classes and then frozen, so heads trained in different stages develop misaligned probability scales and a frozen head can fire confidently on a similar class learned later, turning a horse into a 'cow'. The second, noisy semantics, comes from pseudo-labels used to carve the evolving background into pure background and unknown objects; these labels are unreliable yet are learned with the same strategy as clean labels. IPSeg's first mechanism, image posterior guidance, rests on the Fact that an image's class set is the same whether read at image level or pixel level ($C_I = C_P$): an incrementally trained image branch $\psi$ predicts class-presence probabilities over all seen classes, and the final pixel score is the product of that image posterior with the pixel-wise sigmoid scores (Eq. 3), with a fixed background compensation $\alpha_{BC}=0.9$. The second mechanism, permanent-temporary semantics decoupling, gives a permanently trained branch $\phi_p$ the pure background and unknown foreground, gives each per-task head $\phi_t$ its target classes plus 'other foreground', trains them with different labels (Eq. 5), and discounts spurious 'other foreground' votes at inference (Eq. 9, $\alpha_{NF}=0.4$). On Pascal VOC 2012 and ADE20K the method reports the best results among replay-based competitors in every tested scenario, with the largest advantages in the long-term scenarios VOC 10-1 and VOC 2-2.

Load-bearing premise

The correction's value rests on the image-level posterior branch being more accurate and better calibrated than the pixel-level heads it rescales; if that branch misses a class that is present but small or rare, the multiplication zeroes out every pixel of that class, and the branch was trained partly on pseudo-labels generated by the very heads it is meant to correct.

Editorial extensions

If this is right

Long-horizon scenarios are where the claim bites: on VOC 10-1 and 2-2 with Swin-B the method reports gains of 4.9 and 6.4 mIoU points over the second-best replay method CoinSeg-M, and a 24.8-point gain over SSUL-M on new classes in 2-2.
Separating 'is this class present?' (image posterior) from 'where is it?' (pixel heads) splits the forgetting problem: the appendix's image-level accuracy shows the posterior branch degrades slower than the pixel branch, and the product of the two is steadier than the pixel branch alone.
The decoupling concentrates its benefit on exactly the noisy region: in the VOC 15-1 ablation the new foreground classes gain +5.5 mIoU and background +1.9, while base classes gain less, so the scheme mainly rescues classes learned with pseudo-labeled surroundings.
The memory buffer is load-bearing but not indispensable: the data-free variant already beats prior data-free methods, while the replay version adds several more points (e.g., 65.5 to 72.4 in VOC 2-2), and the authors note the buffer's privacy cost as the main limitation.
The overhead is bounded: roughly 29.7M extra parameters and 27.3 vs 33.7 FPS at inference on VOC 15-1, while storing only salient masks instead of full annotations reduces buffer storage to 1/8.

Reading between the lines

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

Because the Fact is not segmentation-specific, the same product correction could transfer to other pixel-wise continual settings — incremental panoptic segmentation, detection, or open-vocabulary segmentation — as long as an image-level presence predictor stays better calibrated than the per-pixel heads.
The method's tolerance for rare or small objects is untested: a posterior that misses a small cow zeroes out every cow pixel, and the paper's ablations report only aggregate image-level accuracy, so a stress test on images with small or uncommon targets would reveal whether the correction trades precision for recall.
The pseudo-label loop is self-referential — the posterior is trained partly on labels produced by the same heads it corrects — so deliberately corrupting early heads and measuring how much error propagates into the posterior would isolate how much of the gain is genuinely a new signal rather than the model agreeing with itself.
The insensitivity to $\alpha_{BC}$ between 0.7 and 1.0 hints that the mechanism acts as a class-presence gate rather than a precise probability estimate, which suggests a simpler calibration rule might capture much of the benefit without an extra branch.

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

Short version: this is a solid empirical paper that deserves a real referee. The central idea—use an incrementally trained image-level posterior to rescale pixel logits from frozen task heads—is new in CISS and is backed by a thorough experimental campaign. The permanent/temporary branch split and the storage-efficient buffer are sensible engineering choices. I would send it out.

What's genuinely new: the image posterior branch trained with ground-truth plus pseudo-labels, and the multiplicative correction at inference. The paper also ships code and gives cost analysis (params, FLOPs, FPS), which is more than most papers in this area do.

Soft spots, in order. First, the abstract's '24.8% improvement' does not match Table 1; I can't reproduce that number from the reported mIoUs. It should be corrected. Second, Table 2 has an implausible value: SSUL-M† on ADE 100-50 with 101-150 = 48.0 while the same row's all = 33.7. Looks like a typo. Third, the image posterior branch is trained partly on pseudo-labels from the same heads it corrects, so its independent accuracy is not established. The appendix Tables 6/7 give image-level accuracy and show the IP branch drifts less than the pixel branch, which is evidence in the right direction, but the paper never reports IP branch per-class accuracy or failure cases for rare/small objects. The concern that Eq. 3 can suppress a true class when the IP branch misses it is real; the Full-GT vs Pseudo ablation in Table 3 quantifies a 2.4-point cost on VOC 2-2, which is a material fraction of the reported 6.4-point gain over CoinSeg-M. Still, even the Full-GT upper bound beats CoinSeg-M by a wide margin, so the mechanism is not an artifact—it's under-analyzed, not broken. Fourth, no error bars or multiple seeds anywhere, which is common in this literature but worth noting.

Verdict: conditional accept in spirit; needs a number-checking pass and one additional analysis of the IP branch's reliability per class. The citation pattern looks fine. The self-identified limitation on memory buffer is accurately stated.

Who it's for: anyone working on class-incremental segmentation or replay-based continual learning. It's a useful baseline and a reproducible one. I'd bring it to reading group and would cite it.

Referee Report

4 major / 6 minor

Summary. The paper addresses class-incremental semantic segmentation (CISS), where the meaning of background and foreground pixels changes across tasks. It identifies two sources of semantic drift: separate optimization of per-task heads, which produces misaligned probability scales, and noisy pseudo-labels. The proposed IPSeg method adds an image-level posterior branch that is trained incrementally and multiplies pixel-wise probabilities by image-posterior probabilities at inference (Eq. 3). It also decouples learning into a permanent branch for stable background/unknown semantics and temporary task heads for target classes, with a noise-filtering trick. Experiments on Pascal VOC 2012 and ADE20K with ResNet-101 and Swin-B backbones report consistent improvements over existing replay and data-free methods, with the largest gains in long-horizon scenarios such as VOC 10-1 and 2-2.

Significance. If the reported results hold, IPSeg would be a useful contribution to CISS: it attacks a real and under-studied problem (probability-scale misalignment across frozen task heads), provides a concrete mechanism (image-posterior guidance), and supports the claim with extensive experiments over two datasets, two backbones, multiple incremental scenarios, ablations, hyperparameter sensitivity, cost analysis, and qualitative results. The code is released, and the paper includes machine-checkable tables and appendices that make the empirical claims reproducible. The main scientific value is the long-horizon gain on VOC 2-2 and 10-1, where prior methods degrade sharply. However, the central mechanism of Eq. (3) has a robustness risk that is not yet quantified: an image-level false negative can suppress all pixels of a true class, and the image-posterior branch is trained with pseudo-labels from the very heads it is supposed to correct. The paper's aggregate accuracy tables do not resolve this concern.

major comments (4)

[§3.3, Eq. (3)] The inference-time rectification multiplies every pixel-level sigmoid by the image-level sigmoid ψ(c|x). If ψ returns a low probability for a class that is actually present but small or rare in the image, every true-positive pixel of that class is suppressed, since the product is near zero. The paper's aggregate image-level accuracy in Appendix Tables 6 and 7 does not rule this out, because it reports only overall percentages, not per-class recall or the frequency of hard suppression. This concern is load-bearing because the headline gains in VOC 10-1 and 2-2 are attributed to the image-posterior branch. Please report per-class false-negative rates of ψ and per-class Pixel-vs-Pixel+IP mIoU, or change Eq. (3) to a less aggressive correction (e.g., a floor or a weighted combination) and re-evaluate.
[§3.3, Eq. (2)] The image-posterior branch is trained with labels Y~ = Y ∪ Y~_{φ_{1:t-1}}, where the pseudo-labels come from previous task heads that suffer from the separate-optimization problem the method tries to fix. The branch can therefore inherit systematic confusions such as cow/horse. Table 3 quantifies one consequence: replacing pseudo-labels with full ground-truth image labels raises VOC 2-2 from 72.4 to 74.8, which is 2.4 points of a total 6.4-point gain over CoinSeg-M. The paper does not measure the accuracy of ψ itself on old classes, nor the noise level of the pseudo-labels. Please add an evaluation of ψ's per-class precision/recall with and without pseudo-label supervision, and an ablation using confidence-filtered pseudo-labels.
[Abstract and §1 vs. Table 1] The claim of '24.8% improvement in VOC 2-2' is not supported by the table. With Swin-B, IPSeg's 72.4 mIoU is 23.0 points above SSUL-M (49.4) and 6.4 points above CoinSeg-M (66.0); with ResNet-101 it is 13.6 points above SSUL-M and 9.0 points above MicroSeg-M. The reported number appears inconsistent with the paper's own results. Please correct the stated improvement and specify the baseline, or remove the quantitative claim from the abstract.
[§3.3, 'Fact'] The stated Fact that the image-level class domain CI equals the pixel-level class domain CP is definitionally true when image labels are derived from the set of pixel classes present. It does not by itself justify the inference mechanism, which also assumes that the learned image classifier ψ is more reliable than the pixel heads it corrects. This is an extra assumption that is only indirectly supported by Appendix Tables 6–7. Please state the assumption explicitly and test it per class, including classes that are small or rare in the validation set.

minor comments (6)

[§3.4, Eq. (5)] The permanent branch is described as learning permanent semantics, but the unknown-foreground class c'_u shrinks as tasks progress. Consider calling it the 'stable' or 'continuously updated' branch, or explicitly define what 'permanent' means in terms of training schedule rather than semantic content.
[Figure 2] The green inference data flow is not immediately clear. Please clarify that the image-posterior branch is used only at inference time for Eq. (3), while the segmentation heads also provide pseudo-labels during training.
[Table 11(b)] The text in Section A.3 refers to 'αBR' but the table and Eq. (3) use αBC; please correct the typo.
[Table 8] The sentence saying IPSeg has 29.72M parameters more than SSUL 'due to the additional image posterior branch' should also mention the permanent branch, since both are introduced by the method. Reporting per-epoch training time would also make the cost comparison more actionable.
[Conclusions and Limitations] The Limitations section only mentions the memory-buffer requirement. Given the false-negative risk in Eq. (3), please add a sentence acknowledging that the image-posterior correction can suppress true classes when the image-level branch is uncertain.
[Throughout] The capitalization of the method name is inconsistent ('IPSeg' vs. 'IPseg'). Please make it uniform.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: IPSeg's reported gains are empirical results evaluated against external benchmarks, and the image-posterior pseudo-label dependency is a self-training design choice, not a derivation that reduces to its inputs by construction.

full rationale

IPSeg's central claim is that image-posterior guidance plus permanent/temporary semantics decoupling improves class-incremental semantic segmentation. The derivation chain is: (1) the 'Fact' that image-level and pixel-level class domains coincide, which is a motivation rather than a derived prediction; (2) training the image posterior branch psi with Eq. 2, using ground-truth labels unioned with image-level pseudo-labels produced by previous segmentation heads; (3) rectifying pixel predictions by element-wise multiplication in Eq. 3; and (4) empirical evaluation on Pascal VOC and ADE20K. The only potentially circular-looking dependency is that psi is trained partly on pseudo-labels from the very heads phi_{1:t-1} that it later corrects. However, this is a standard self-training/bootstrap procedure, not a mathematical equivalence: psi is a distinct classifier with pooling and MLP parameters trained by a BCE objective, and Eq. 3's multiplication does not force any particular outcome. The paper directly tests this dependency in Table 3, where replacing pseudo-labels with full ground-truth image labels changes VOC 2-2 mIoU from 72.4 to 74.8, showing the result is not identical to its inputs by construction; if the pseudo-label loop were circular in the sense of merely returning its input, the Full-GT and Pseudo rows would coincide. The ablation also shows that pseudo-labels improve over partial ground truth, so the design has independent empirical content. The self-citations to MicroSeg and CoinSeg (Zhang et al., 2022b, 2023, which share author Yunchao Wei) are used as baselines and as sources of standard techniques such as pseudo-labeling and saliency maps; they do not carry the argument's validity, and the comparisons are against externally published or reproduced results. The skeptical concern that image-level misses can suppress rare or small objects is a legitimate robustness risk, and the paper does not fully measure the image posterior branch's per-class reliability, but that is a correctness or evaluation issue, not a circularity of the derivation chain. Overall, no load-bearing step reduces to its own inputs, so the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 2 invented entities

The core contribution rests on three tuned hyperparameters, two dataset-dependent assumptions about image-level and pixel-level consistency, and two new architectural components. The hyperparameters are small in number and the ablations show moderate sensitivity, but the correctness of the mechanism depends on the saliency detector and self-generated pseudo-labels, which are externally unverified. The paper is honest in the Limitations section about the memory buffer dependency.

free parameters (3)

background compensation alpha_BC = 0.9
Introduced in Eq. 3 to compensate for the lack of background posterior probability from the image posterior branch, tuned by ablation in Table 11(b). The final results depend on this value.
noise filtering factor alpha_NF = 0.4
Introduced in Eq. 9 to downweight predictions when the dummy foreground cf wins, tuned by ablation in Table 11(a). It is a post-hoc heuristic that affects the reported scores.
loss weights lambda_1, lambda_2 = 0.5, 0.5
Trade-off weights for the temporary and permanent branch losses in Eq. 7, tuned by the ablation in Table 10. The default values produce the best reported result.

assumptions (4)

domain assumption The image-level class domain and pixel-level class domain of any image coincide (Fact, Section 3.3).
This is the conceptual foundation of the image posterior guidance. It is presented as a fact but is a modeling assumption that holds only if the image-level branch is accurate enough.
domain assumption The semantic content of an image in task t divides into past classes, current classes, unknown foreground, and pure background (Observation, Section 3.4).
This decomposition underpins the permanent-temporary decoupling. It assumes a clean separation that the pseudo-labeling pipeline must deliver.
domain assumption The saliency detector S reliably locates unknown foreground objects.
Eq. 5 uses S(x_i)=1 to decide whether background pixels are unknown foreground c'_u or pure background c'_b. Errors in the saliency map propagate directly into the pseudo-labels.
domain assumption The frozen previous task heads provide reliable enough image-level pseudo-labels for training the image posterior branch.
Eq. 2 mixes ground-truth labels with pseudo-labels from previous heads. If these pseudo-labels are systematically wrong, the image posterior branch inherits the error it is designed to fix. Table 3 shows pseudo-labels outperform partial ground truth, but the assumption remains load-bearing.

invented entities (2)

Image posterior branch psi
purpose: Predicts image-level class presence and rectifies pixel-wise predictions via element-wise multiplication in Eq. 3.
This is a new model component. Its benefit is validated only through the paper's own experiments; there is no external benchmark showing that image posterior guidance helps incremental segmentation.
Permanent branch phi_p and temporary branches phi_t
purpose: Decouple stable background/unknown semantics from transient foreground semantics.
Architectural decomposition validated only on the paper's chosen benchmarks. No external evidence that permanent-temporary decoupling generalizes.

how reviews work

0 comments

Cite this review

Pith. "Pith review of IPSeg: Image Posterior Mitigates Semantic Drift in Class-Incremental Segmentation." pith.science (2026). https://pith.science/paper/KPGFLORI

@misc{pith2026250204870,
  author       = {Pith},
  title        = {Pith review of: IPSeg: Image Posterior Mitigates Semantic Drift in Class-Incremental Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KPGFLORI}},
  note         = {Machine review of arXiv:2502.04870}
}

read the original abstract

Class incremental learning aims to enable models to learn from sequential, non-stationary data streams across different tasks without catastrophic forgetting. In class incremental semantic segmentation (CISS), the semantic content of image pixels evolves over incremental phases, known as semantic drift. In this work, we identify two critical challenges in CISS that contribute to semantic drift and degrade performance. First, we highlight the issue of separate optimization, where different parts of the model are optimized in distinct incremental stages, leading to misaligned probability scales. Second, we identify noisy semantics arising from inappropriate pseudo-labeling, which results in sub-optimal results. To address these challenges, we propose a novel and effective approach, Image Posterior and Semantics Decoupling for Segmentation (IPSeg). IPSeg introduces two key mechanisms: (1) leveraging image posterior probabilities to align optimization across stages and mitigate the effects of separate optimization, and (2) employing semantics decoupling to handle noisy semantics and tailor learning strategies for different semantics. Extensive experiments on the Pascal VOC 2012 and ADE20K datasets demonstrate that IPSeg achieves superior performance compared to state-of-the-art methods, particularly in challenging long-term incremental scenarios.

Figures

Figures reproduced from arXiv: 2502.04870 by the authors.

**Figure 1.** (a) Due to the existence of separate optimization, the previous method SSUL-M misclassifies a “horse” as a “cow” with higher logit scores when learning “horse” following “cow”. While our IPSeg leverages image posterior (IP) guidance to produce accurate predictions on these two similar-look classes. The “logit scores” refer to pixel-wise prediction, and the image posterior refers to our introduced image-wise predicti… view at source ↗

**Figure 2.** Overall architecture of our proposed IPSeg, mainly composed of image posterior and permanent-temporary [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** (a) The overall performance of different methods on Pascal VOC 2012 under 4 scenarios, (b) mIoU visualization [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: The visualization of separate optimization. [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]

**Figure 5.** Figure 5: The probability distributions for SSUL-M, IPSeg, and Joint-Training (Joint) in the regions of incorrect predictions. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

**Figure 6.** Figure 6: Semantics decoupling strategy sofa cow image step 0 step 2 step 4 step 6 step 8 step 9 GT 8 chair 9 cow 10 table 12 horse 14 person 17 sofa SSUL-M CoinSeg-M IPSeg (Ours) table table cow sofa sofa sofa cow cow cow cow [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

**Figure 7.** Figure 7: Qualitative analysis of IPSeg on Pascal VOC 2012. Texts and bounding boxes in white indicate incorrect class [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

**Figure 8.** Figure 8: Qualitative results on Pascal VOC 2012 dataset with the 15-1 scenario. [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]

**Figure 9.** Figure 9: Qualitative results on ADE20K dataset with the 100-10 scenario. [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 12 canonical work pages

[2]

10” and “17

0.40510.3725 00.10.20.30.40.50.60.70.80.91 01234567891011121314151617181920 Probability Distribution Class Index SSUL-M 0.1022 0.8017 00.10.20.30.40.50.60.70.80.91 01234567891011121314151617181920Class Index Joint 0.2541 0.5162 00.10.20.30.40.50.60.70.80.91 01234567891011121314151617181920Class Index IPSeg Image Prediction GT Figure 5: The probability dis...

work page arXiv
[3]

Imagenet: A large-scale hierarchical image database

Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pp. 248–255. Ieee,

2009
[5]

Gurbuz, M. B. and Dovrolis, C. Nispa: Neuro-inspired stability-plasticity adaptation for continual learning in sparse networks. arXiv preprint arXiv:2206.09117,

work page Pith review arXiv
[7]

Recent advances of continual learning in computer vision: An overview

Qu, H., Rahmani, H., Xu, L., Williams, B., and Liu, J. Recent advances of continual learning in computer vision: An overview. arXiv preprint arXiv:2109.11369,

arXiv
[8]

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,

2001
[9]

We select five examples to illustrate the model’s ability to predict various classes as the learning step increases. Row 1 shows the performance of predicting the new class ”ship” in step 1, where the model effectively recognizes both the old class ”sky” and the new class ”ship.” Similarly, in rows 2-5, for the newly introduced classes (tent, oven, screen...

work page 2012
[10]

Early preparation pays off: New classifier pre-tuning for class incremental semantic segmentation

Xie, Z., Lu, H., Xiao, J.-w., Wang, E., Zhang, L., and Liu, X. Early preparation pays off: New classifier pre-tuning for class incremental semantic segmentation. arXiv preprint arXiv:2407.14142,

arXiv
[12]

Adaptive Prototype Replay for Class Incremental Semantic Segmentation

Zhu, G., Wu, D., Gao, C., Wang, R., Yang, W., and Sang, N. Adaptive prototype replay for class incremental semantic segmentation. arXiv preprint arXiv:2412.12669,

work page Pith review arXiv

Show all 18 references

[13]

cow” and “horse

11 IPSeg: Image Posterior Mitigates Semantic Drift in Class-Incremental Segmentation A. Appendix A.1. Symbols and Explanations Table 5 provides key symbols used in our paper along with their explanations to facilitate a better understanding. Table 5: Symbols and explanations S...

2012
[15]

We set image size=512x512, epochs=50, and batch size=16 in training and image size=512x512 for inference test

We test and report the results of IPSeg, SSUL-M and CoinSeg-M with Swin-B on the VOC 15-1 setting. We set image size=512x512, epochs=50, and batch size=16 in training and image size=512x512 for inference test. All results are run on RTX 3090 GPU. • Model Parameters: Using the ...
[17]

Additionally, similar to the results in the overlap setting, IPSeg exhibits a strong ability to learn new classes while retaining knowledge of the old classes

The results indicate that IPSeg consistently achieves the best performance compared to state-of-the-art methods. Additionally, similar to the results in the overlap setting, IPSeg exhibits a strong ability to learn new classes while retaining knowledge of the old classes. Spec...
[2009]

Podnet: Pooled outputs distillation for small-tasks incremental learning

Douillard, A., Cord, M., Ollion, C., Robert, T., and Valle, E. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Computer vision–ECCV 2020: 16th European conference, Glasgow, UK, August 23–28, 2020, proceedings, part XX 16 , pp. 86–102. Springer,

2020
[2012]

cow” and “horse

IPSeg demonstrates superior performance in various incremental learning tasks, including standard tasks with a large number of initial classes (e.g., 15-5 and 15-1) and long-range tasks with fewer initial classes (e.g., 10-1 and 2-2). Notably, in the 2-2 task, the mIoU for “co...

2012
[2017]

L., Yang, Q., and Li, L

Silver, D. L., Yang, Q., and Li, L. Lifelong machine learning systems: Beyond learning algorithms. In 2013 AAAI spring symposium series,

2013
[2020]

and Liu, B

Ke, Z. and Liu, B. Continual learning of natural lan- guage processing tasks: A survey. arXiv preprint arXiv:2211.12701,

arXiv
[2021]

Diffusepast: Diffusion-based generative replay for class incremental semantic segmentation.arXiv preprint arXiv:2308.01127,

Chen, J., Wang, Y ., Wang, P., Chen, X., Zhang, Z., Lei, Z., and Li, Q. Diffusepast: Diffusion-based generative replay for class incremental semantic segmentation.arXiv preprint arXiv:2308.01127,

arXiv
[2023]

Rethinking atrous convolution for semantic image seg- mentation

Chen, L.-C., Papandreou, G., Schroff, F., and Adam, H. Rethinking atrous convolution for semantic image seg- mentation. arXiv preprint arXiv:1706.05587,

arXiv
[2024]

and Zhao, D

Yuan, B. and Zhao, D. A survey on continual semantic segmentation: Theory, challenge, method and application. arXiv preprint arXiv:2310.14277,

arXiv

Pith tools

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

[1] [2]

10” and “17

0.40510.3725 00.10.20.30.40.50.60.70.80.91 01234567891011121314151617181920 Probability Distribution Class Index SSUL-M 0.1022 0.8017 00.10.20.30.40.50.60.70.80.91 01234567891011121314151617181920Class Index Joint 0.2541 0.5162 00.10.20.30.40.50.60.70.80.91 01234567891011121314151617181920Class Index IPSeg Image Prediction GT Figure 5: The probability dis...

work page arXiv

[2] [3]

Imagenet: A large-scale hierarchical image database

Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pp. 248–255. Ieee,

2009

[3] [5]

Gurbuz, M. B. and Dovrolis, C. Nispa: Neuro-inspired stability-plasticity adaptation for continual learning in sparse networks. arXiv preprint arXiv:2206.09117,

work page Pith review arXiv

[4] [7]

Recent advances of continual learning in computer vision: An overview

Qu, H., Rahmani, H., Xu, L., Williams, B., and Liu, J. Recent advances of continual learning in computer vision: An overview. arXiv preprint arXiv:2109.11369,

arXiv

[5] [8]

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,

2001

[6] [9]

We select five examples to illustrate the model’s ability to predict various classes as the learning step increases. Row 1 shows the performance of predicting the new class ”ship” in step 1, where the model effectively recognizes both the old class ”sky” and the new class ”ship.” Similarly, in rows 2-5, for the newly introduced classes (tent, oven, screen...

work page 2012

[7] [10]

Early preparation pays off: New classifier pre-tuning for class incremental semantic segmentation

Xie, Z., Lu, H., Xiao, J.-w., Wang, E., Zhang, L., and Liu, X. Early preparation pays off: New classifier pre-tuning for class incremental semantic segmentation. arXiv preprint arXiv:2407.14142,

arXiv

[8] [12]

Adaptive Prototype Replay for Class Incremental Semantic Segmentation

Zhu, G., Wu, D., Gao, C., Wang, R., Yang, W., and Sang, N. Adaptive prototype replay for class incremental semantic segmentation. arXiv preprint arXiv:2412.12669,

work page Pith review arXiv