Fairness Continual Learning Approach to Semantic Scene Understanding in Open-World Environments

Continual semantic segmentation aims to learn new classes while maintaining the information from the previous classes. Although prior studies have shown impressive progress in recent years, the fairness concern in the continual semantic segmentation needs to be better addressed. Meanwhile, fairness is one of the most vital factors in deploying the deep learning model, especially in human-related or safety applications. In this paper, we present a novel Fairness Continual Learning approach to the semantic segmentation problem. In particular, under the fairness objective, a new fairness continual learning framework is proposed based on class distributions. Then, a novel Prototypical Contrastive Clustering loss is proposed to address the significant challenges in continual learning, i.e., catastrophic forgetting and background shift. Our proposed loss has also been proven as a novel, generalized learning paradigm of knowledge distillation commonly used in continual learning. Moreover, the proposed Conditional Structural Consistency loss further regularized the structural constraint of the predicted segmentation. Our proposed approach has achieved State-of-the-Art performance on three standard scene understanding benchmarks, i.e., ADE20K, Cityscapes, and Pascal VOC, and promoted the fairness of the segmentation model.

Paper

Similar papers

Peer review

Reviewer 9fXu6/10 · confidence 5/52023-06-27

Summary

This paper proposes a fairness continual learning approach in semantic segmentation. Derived from the therotical analysis of the effect of class distribution, a new learning paradigm that is consisted of three main objectives: the task-specific objective to handle the catastrophic forgetting problem, the fairness objective that maintains the fairness predictions over classes, and the conditional structural constraint that imposes the consistency of the segmentation predictions. New SOTA results are attained on the standard benchmarks.

Strengths

1. Deriving the learning objective through analysis of the effect of class distribution is novel. 2. Strong performances are achieved on several standard benchmarks.

Weaknesses

- The title is misleading. In my understanding, the open-world environments means lacking of labels. While the proposed method can hardly work in unsupervised setting. - Some explanation through mathmatical derivation is not rigorous such as eq. 5 and eq. 12 and 13. More explanation either in the main paper or supplimentary material needed. Detail illustration can be found in the questions. - Some details are not clear. How the fairness objective $L_{class}$ is calculated during training is not clear and lacking a final integrated training objective. Whether there is hyperparameters between losses is not mentioned. - According to the ablation study, fairness improvement brought by the fairness objective is minor if measured in STD of IoU over all classes. Better measurement of the fairness is needed. For example, variation of the performance on rare classes and dominant classes with and without fairness loss can be compared in a longer learning steps.

Questions

1. Eq. 5 is hard to understand and I think it could not be intuitively achieved. 2. The explanation that the prototypical contrastive clustering loss is an upper bound of the knowledge distillation approach is not rigorous. In eq. (12) , the $D(p_c, f^{t-1})$ cannot directly ignored as a constant number. When $c$ is a new class from $C^t$, the $p_c$ is changing during learning and $D(p_c, f^{t-1})$ will change. Therefore, $O(D(f^t, p_c)$ is not a tight boundary. Minimizing $L_{cluster}$ is not strictly equals to the minimizing of $L_distill$. From another perspective, although prototypical constrative clustering loss don't need two forward pass of the segmentation model, it have to store the prototypical vector $p_c$ using additional memory. How to calculate the $p_c$ is also critical in achieving good performance. 3. What is the final integrated training loss? Is there any balance weights between different losses? 4. What is the result in the first line of DeepLab-V2 and SegFormer in Table 1? 5. Im my understanding, the conditional structural consistency mitigate the prediction gap among neighbour pixels. However, it will also smooth the edge pixels of an instance. According to the ablation study, it also improves the fairness of classes. Could the author give more explanation on it? 6. Minor problems: when writting the related works, it better to use the surname other than the name. For example, Ziwei et al. should be Liu et al in line 93. In eq. 5, the $y_{i,h}$ should be $y_{i,j}$ in the left term?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Authors have pointed out the limitations of the proposed method and some possible improvements.

Reviewer 2Nqo6/10 · confidence 3/52023-07-05

Summary

This paper proposes a Fairness Continual Learning approach to address the continual semantic segmentation problem. Specifically, the fairness continual learning framework is proposed for fairness objective, while a Prototypical Contrastive Clustering loss for reducing catastrophic forgetting. When compared to prior work on continual semantic segmentation, FairCL is shown to improve Final mIoU significantly.

Strengths

1. This paper analyzes several problems that need to be solved in continuous learning in detail, and clearly presents the loss function corresponding to the three objectives from the perspective of formula derivation. 2. Extensive ablation studies are performed which provides insight into the challenges of continual semantic segmentation problem.

Weaknesses

1. The organization of method is not very clear. It's difficult to follow what the final loss function is and how each part of loss is combined. 2. About the Prototypical Contrastive Clustering Loss, I wonder whether pulling all features of unknown classes into one cluster is reasonable. Will learning similar features from categories with significant differences affect subsequent model training? 3. How is the margin $\triangle$ selected? 4. In Sec.3.2, 'a pseudo label where its assigned label is computed based on the closest cluster'. How is the pseudo label computed and utilized? 5. The compare in experiments about network backbone is unfair. The experiment provides results on two different backbones, but neither backbone was aligned with the previous method such as Deeplab-v3 architecture with ResNet-101 in RCIL.

Questions

I am looking forward to the authors clarifying the concerns listed in the section above: 1. The final loss function and the combination form of each part; 2. More detail explanation of the Prototypical Contrastive Clustering Loss; 3. Comparison in experiment results.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

Both limitations and potential societal impacts are addressed in the paper. However, it might help to expand the paragraph to discuss some of the concerns listed.

Reviewer ReY55/10 · confidence 4/52023-07-05

Summary

The paper explores the continual semantic segmentation problem and highlights the lack of fairness in existing methods. It proposes a novel fairness objective based on class distributions to address the bias on the majority group classes. The paper proposes a novel prototype contrastive clustering loss to address the significant challenges of catastrophic forgetting as well as for handling unknown classes to solve the background shift problem. Furthermore, a conditional structural consistency loss is proposed to constrain the structure of the predicted segmentation. The method has achieved improvements over the existing benchmarks on ADE20K, Cityscapes and Pascal VOC datasets.

Strengths

1. The paper is well written and the motivation is good. The problem formulation is nicely done. 2. The paper points out the issue of maintaining fairness in continual segmentation systems. It is a very significant contribution to the field since all existing methods primarily focussed on reducing forgetting and addressing the background shift.  The fairness problem is well-defined in the paper along with the fairness metric which is used to obtain the error-rate among the classes. 3. The novel prototype contrastive clustering loss represents the semantic segmentation pixels for different classes in the latent space and groups the features of unknown classes into the same cluster. In addition to handling unlabelled classes, it is interesting to see that the clustering loss also serves as an upper bound of the knowledge distillation approach and thus reduces the forgetting of old classes. 4. The proposed method FairCL seems to be quite effective on several settings across multiple datasets with different backbones.

Weaknesses

1. Although the fairness loss is improving the mIoU values for different settings on ADE20K, it is not clear if the improvement in mIoU is coming from the minority classes. While the main motivation of the paper is to ensure fairness among the classes, the ablation experiment does not completely justify this. The individual contribution of the fairness loss (L_class) should be stated since Fairness in Continual Learning is the main claim of the authors. Further analysis of the improvements in mIoU values separately for majority and minority classes on at least two datasets will be more convincing. Also, the fairness loss may be evaluated in a non-incremental setting (let's say for the first step only), and then the minority class mIoU values should be better; otherwise, the performance is also affected by forgetting, and the contribution of the fairness loss is not clear. 2. The motivation for learning conditional structural consistency loss is not clear and should be discussed to give more intuition. Also, some qualitative analysis of prediction consistency will be good to show the effectiveness of the structural consistency loss. 3. Results from Pascal VOC are not convincing. For the 10-1 setting, the forgetting is quite high compared to methods like RCIL. Overall, the improvement looks very insignificant. Experiments on more settings of Pascal VOC similar to previous papers like PLOP and RCIL will be good. Also, the comparison of the SegFormer backbone with the other existing methods using DeepLabV2 is not fair. If there is any specific reason for the method not being so effective for Pascal VOC (as it looks effective for ADE20K and Cityscapes), it should be analyzed.

Questions

1. Inconsistent results: The results for the ADE20K 100-50 setting are different in table 1 and table 3. In table 1 for ablation, the overall mIoU is 34.64 for DeepLab-V2 backbone, while in table 3, the overall mIoU is 36.99. The mIoU values are the same for the SegFormer backbone though. This should be clarified since the ablation experiment is claiming the contributions of the different losses. 2. While it is good to see the results with both DeepLab-V2 and SegFormer backbones, the backbone networks of the compared methods should be mentioned for more clarity and fair comparison. 3. The comparison does not include a well performing baseline method - SSUL (from NeurIPS 21). If there is any specific reason for not comparing with SSUL, it should be discussed in the paper. 4. See weaknesses 1, 2, 3.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

/

Reviewer jVAT7/10 · confidence 4/52023-07-24

Summary

The paper focuses on the semantic segmentation problem in the context of continual learning. In the past, the emphasis was mainly on addressing catastrophic forgetting and background shift issues. However, this paper goes beyond that and investigates the balance problem between different object categories. It introduces novel approaches for prototype contrastive learning and structure-consistency loss functions, which result in a new SOTA performance. The contributions of the paper can be summarized into three main parts: i. It introduces a novel evaluation metric to measure the model's balance across different classes. ii. It proposes a new contrastive learning loss to address the issue of forgetting in continual learning. iii. The paper conducts extensive experiments to demonstrate the effectiveness and rationality of the proposed methods.

Strengths

+ The paper considers multiple perspectives to improve fairness among classes of the model. + The paper designs various loss functions to enhance the segmentation performance of the model under continual learning. + The supplementary materials of the paper include an excellent presentation that makes it easier for readers to understand the superiority of the model.

Weaknesses

+ The paper lacks experiments specifically focused on the upper bound design for continual learning. + The proposed inter-class fairness in the paper seems to be addressing the semantic segmentation problem rather than the main focus of continuous learning. However, this issue has already been well discussed in [1]. [1] Truong, T. D., Le, N., Raj, B., Cothren, J., & Luu, K. (2023). Fredom: Fairness domain adaptation approach to semantic scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 19988-19997).

Questions

As far as I know, the current research on semantic segmentation is primarily conducted using DeepLab v3 as the framework. Why did the paper choose to experiment with DeepLab v2 instead?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The paper discusses its own limitations in the experimental section. The paper does not have any negative social impact.

Reviewer ReY52023-08-18

Response for the authors

Thank you for your detailed responses, which addresses most of my concerns. I will maintain my score.

Authorsrebuttal2023-08-18

Thanks For Rating

Dear Reviewer ReY5, Thank you very much for your positive feedback and rating. We are committed to updating our paper based on the reviewers' suggestions. Thank you very much, Authors

Reviewer jVAT2023-08-18

Thank the authors for their responses. I don't have any more questions. I hope the mentioned issues can be addressed in the future version, and I will raise my score to 7.

Authorsrebuttal2023-08-18

Thanks For Raising Rating

Dear Reviewer jVAT, Thank you very much for your positive feedback and raising your rating. We are committed to updating our paper based on the reviewers' suggestions. Thank you very much, Authors

Reviewer 9fXu2023-08-21

Thanks for the detailed response. I will keep my rating.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC