SSA-Seg: Semantic and Spatial Adaptive Pixel-level Classifier for Semantic Segmentation

Vanilla pixel-level classifiers for semantic segmentation are based on a certain paradigm, involving the inner product of fixed prototypes obtained from the training set and pixel features in the test image. This approach, however, encounters significant limitations, \ie, feature deviation in the semantic domain and information loss in the spatial domain. The former struggles with large intra-class variance among pixel features from different images, while the latter fails to utilize the structured information of semantic objects effectively. This leads to blurred mask boundaries as well as a deficiency of fine-grained recognition capability. In this paper, we propose a novel Semantic and Spatial Adaptive Classifier (SSA-Seg) to address the above challenges. Specifically, we employ the coarse masks obtained from the fixed prototypes as a guide to adjust the fixed prototype towards the center of the semantic and spatial domains in the test image. The adapted prototypes in semantic and spatial domains are then simultaneously considered to accomplish classification decisions. In addition, we propose an online multi-domain distillation learning strategy to improve the adaption process. Experimental results on three publicly available benchmarks show that the proposed SSA-Seg significantly improves the segmentation performance of the baseline models with only a minimal increase in computational cost. Code is available at https://github.com/xwmaxwma/SSA-Seg.

Paper

References (63)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer Dw1G5/10 · confidence 3/52024-07-12

Summary

In this paper, it analyzes that current pixel-level classifiers for semantic segmentation suffers limitations such as feature deviation in the semantic domain and information loss in the spatial domain. To this end, the authors propose a novel Semantic and Spatial Adaptive (SSA) classifier. Specifically, the authors employ the coarse masks obtained from the fixed prototypes as a guide to adjust the fixed prototype towards the center of the semantic and spatial domains in the test image. In addition, the authors propose an online multi-domain distillation learning strategy to guide the adaption process. Experimental results on three publicly available benchmarks show that the proposed SSA significantly improves the segmentation performance of the baseline models with only a minimal increase in computational cost.

Strengths

The main strengths are as follows: 1. A semantic and spatial adaptive (SSA) classifier is proposed, which facilitates the offset of the fixed prototype towards the center of the semantic domain and the center of the spatial domain of the test image. 2. This paper designs multi-domain knowledge distillation to enhance the primary classifier from different domains. First, the response domain distillation distills the outputs of the two classifiers based on a boundary-aware and category-aware distillation loss, which conveys accurate semantic and structural information in the ground truth. 3. The proposed method significantly improves the segmentation performance of the baseline model with a very small increase in computational cost on three commonly used datasets: ADE20k, PASCAL-Context, and COCO-stuff-10K.

Weaknesses

The main weaknesses are as follows: 1. Why not evaluate the proposed methods on high resolution dataset, like Cityscape. How about the memory usage comparison? what's the input image resolution for training and inference? These are key information need to put in this paper. 2. For the online multi-domain Distillation, in line 199 "we first create a new branch with pixel", what does the "new branch" mean? There is no detail for the "Teacher classifier".

Questions

No. please see weakness part.

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

No. please see weakness part.

Reviewer bEdU2/10 · confidence 5/52024-07-14

Summary

This work primarily focuses on semantic segmentation. Specifically, a semantic and spatial adaptive (SSA) classifier is proposed to address the "feature deviation in the semantic domain and information loss in the spatial domain" issues. The proposed classifier mainly consists of - Semantic Prototype Adaptation, - Spatial Prototype Adaptation, - Online Multi-Domain Distillation Learning. The authors claim that "experimental results on three publicly available benchmarks show that the proposed SSA significantly improves the segmentation performance of the baseline models with only a minimal increase in computational cost."

Strengths

The strengths of this paper can be summarized as follows, - Overall, the author clearly introduces the research motivation and the proposed method. - The author validates the effectiveness of the proposed algorithm across different datasets and segmentation frameworks.

Weaknesses

The weaknesses of the paper are as follows, - The authors claim in the abstract that the code for the proposed algorithm is included in the supplementary materials, but I did not find any related code there. - Some expressions in the paper feel redundant, e.g., "directly classify masks by directly learning query vectors". - In the introduction section, the authors argue that previous mask-level classification models have cumbersome decoders. Therefore, the authors should conduct a comprehensive comparison (params, fps, flops, miou, etc) between their method and these previous methods in the experiments section. - Tables 1 to 3 only provide the FLOPs metric but do not include FPS and parameters, which are also crucial for deploying the model in resource-constrained environments. - Can the authors provide any experiments or references to support the statement "Due to complex backgrounds and varying object distributions, pixel features in the test images tend to have a large intra-class variance with pixel features in the training set"? I think that a single image in Figure 1 can not prove this point. - To be honest, Figure 1 is very confusing to me. For example, what do the gray dots represent? (pixel features belonging to other categories?) Can you provide the model's prediction results for this image? If your visualization is accurate, the prediction should include many pixels of other categories being classified as “door”. In other words, I believe there is a high likelihood that your visualization is unreliable, namely, in the t-SNE visualization results, being close to a semantic prototype does not necessarily mean that the model classifies it as the corresponding semantic category in the feature space. - If the proposed SSA strategy truly facilitates “the offset of the fixed prototype towards the center of the semantic domain and the center of the spatial domain of the test image,” why don't the authors directly visualize this result using t-SNE like Fig. 1(b) instead of presenting a less credible schematic diagram? - I disagree with the content in lines 36 to 40 of the paper. Modeling the relationship between prototypes and pixel features in the spatial domain is incorrect. This means that when classifying pixels, models need to consider which types of objects generally appear at that pixel location. This is obviously an overfitting to the current dataset and has no practical application value. - The authors argue that "pixel features in the test images tend to have a large intra-class variance with pixel features in the training set" and their proposed method can address this issue. If this is indeed the case, the author should apply their method to cross-domain segmentation tasks to better demonstrate its effectiveness. - The content in the Related Works section is too brief. - In line 126, the authors mention that "Based on the above analysis", where is the analysis? - Can the authors provide any experiments or references to support the statement between line 150-151? - The content stated in lines 155 to 157 does not correspond with the formula. - Section 3.2.1 essentially just integrates features similar to class prototypes and does not achieve the author's claimed “facilitating the offset of the fixed prototype towards the center of the semantic domain.” - “utilizing the rich spatial information of the image” should be the responsibility of context modules (such as aspp and ppm) in the decoder head. The premise of section 3.2.2 is fundamentally flawed. - Would the model's performance improve if only “randomly initialized position basis” is used? Can the authors provide related ablation experiments? - If I understand correctly, section 3.2.2 essentially assumes that objects of the same class should all be in a specific region of the image. However, this assumption is unreasonable because, in semantic segmentation, different instances of the same class can be far apart. - To be honest, I believe the methods in sections 3.2.1 and 3.2.2 are not highly related to the problem analyzed by the authors or the expected solutions to that problem. - Would the model's performance improve if section 3.2.3 was replaced with self-distillation? - Due to the existence of $\mathcal{M}_c$, the problem mentioned by the authors in the introduction section still persists in the proposed algorithm. In my opinion, the performance improvement is mainly due to the addition of more parameters and the introduction of more loss functions. While these changes have brought about performance gains, they do not align with the authors' motivation. Due to these weaknesses, I believe this paper should be rejected.

Questions

The main issues with this work have already been listed in the weaknesses box. The presence of numerous unreliable contents and my belief that the premise of the proposed algorithm is unreasonable are the main reasons I am inclined to reject this paper.

Rating

2

Confidence

5

Soundness

2

Presentation

2

Contribution

2

Limitations

The author does not discuss the limitations of their algorithm in the paper. I have provided several suggestions for improvement in the weaknesses box, which I hope will help the authors enhance the quality of this work.

Reviewer ra2D5/10 · confidence 3/52024-07-15

Summary

This paper proposes an adaptive method to improve the semantic segmentation quality. The main idea is to adaptively update the prototypes by using the coarse segmentation masks predicted by the baseline method. Both semantic and spatial prototypes are employed to achieve complementary improvement. Extensive experimental evaluation has been conducted on three public benchmark datasets, showing consistent improvement over different baseline methods.

Strengths

1. The method seems to be novel. The computation overhead of the method is negligible. 2. The results show that the proposed method can consistently improve many segmentation methods, in particular, the methods with lower latencies. 3. Extensive ablation studies are included to show the contributions of different components.

Weaknesses

1. Not every loss term in the Equation 3 is ablated, for example L_c and L_dice 2. Since mask-level classifiers become more and more popular, how could be proposed method be adapted to mask-level methods such as mask2former?

Questions

See weakness

Rating

5

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

Limitations are discussed in the appendix.

Authorsrebuttal2024-08-14

## **Sec3.2.2(SPPA) assumes that objects of the same class should all be in a specific region of the image** ------ Thank you for your response. We have responded to reviewer bEdU in response to your question. However, due to the character limit, it is possible that our answer was not as convincing as it could have been. Here we provide a more detailed answer. - First, we do not assumes that objects of the same class should all be in a specific region of the image in our paper. This comment is unfounded and reviewer bEdU may not have fully understood our method. Although SPPA proposes the concept of spatial domain centers, our spatial domain centers are obtained based on the Conditional Positional Vocoding (CPVT) [1], which represents for relative segments rather than absolute anchor points. Therefore, our spatial domain center models the collection of information about neighboring segments belonging to the same object, rather than an absolute one point. We have a corresponding analysis in lines 314-321 of the paper. In other words, it enables the model to indirectly take into account the semantic features of pixels at neighboring locations when classifying them. More specifically, you can refer to CPVT [1], and we believe you can understand how we avoid modeling absolute spatial anchor points. - Second, many images in ADE20k and COCO-stuff have the same category scattered in different areas. Nevertheless, our classifier SSA can still significantly improve the performance of the model on these datasets. This is because SPPA takes into account the spatial structure information. We have conducted ablation experiments on SPPA in the paper, as shown in Table 3 of the paper. When SPPA and distillation loss are applied, the model accuracy increased by 2.12% mIoU. This validates the effectiveness of spatial domain adaptation. We hope you and AC are not misled by this erroneous comment. ## **Clarifying the motivation** ------ As described in lines 3-8, 29-40 of the paper, our motivation is clear, i.e., the vanilla softmax classifier uses the inner product of pixel features and fixed prototypes to generate segmentation masks, leading to feature deviation in the semantic domain and information loss in the spatial domain problem. Therefore, we introduce semantic prototype adaptation and spatial prototype adaptation, adjust the fixed prototype to the center of the semantic and spatial domains in the test image, and then consider both semantic and spatial domains of the adaptive prototypes to complete the classification decision, which effectively mitigates the above problems. ## **Providing evidences of large intra-class variance with pixel features** ------ We provide more graphic illustrations in Fig.1 of rebuttal. Specially, Fig.1 shows that the t-SNE of some example images, which are randomly selected from the ADE20K dataset. The first row represents the distribution of pixel features in the door class, and the second row represents table class. it can be observed that due to the complex scenarios and varying object distributions, pixel features of the same class tend to exhibit larger intra-class variance when the trained model on the training set is applied to the test set.We will add these graphic illustrations and a more detailed analysis to the revised version of the paper. ## **SSA+Mask2Former** ------ We have further validated your confusion. Specifically, after we embedded SSA into Mask2Former's mask classifier, the mIoU of the model increased from 47.2 to 48.5. This shows that SSA is able to adapt to existing mask classification methods. ## **Conclusion** ------ We have responded to each of the reviewer bEdU's comments. If you have additional questions or find one of our responses unconvincing, please do not hesitate to point it out. We look forward to further discussions with you. [1] Conditional Positional Encodings for Vision Transformers (ICLR 2023)

Authorsrebuttal2024-08-13

Looking forward to further discussions

Dear Reviewers, Thank you for your effort in reviewing our paper. We kindly remind you that the end of the discussion stage is approaching. During the rebuttal period, we made every effort to address your concerns faithfully, including more comparison experiments, more ablation studies, more visual analysis, etc. In particular, we have provided detailed answers to all the questions you have asked, and the codes in the supplement can be used to verify the validity of the SSA. We believe we have effectively addressed most of your concerns and we look forward to your response or any further discussions. Sincerely, Paper3694 Authors

Reviewer ra2D2024-08-13

Thanks for the clarifications which addressed my concerns. However, I am not very confident about my initial rating because I am not in the same field. After reading other reviews, I decided to decrease my rating to be a Borderline accept. In particular, I agree with reviewer bEdU that the writing of this paper could be significantly improved, especially clarifying the motivation and providing evidences of large intra-class variance with pixel features. Although the rebuttal addressed many questions of reviewer bEdU, some weakness points are still valid and make the paper less convincing. For example, reviewer bEdU points out that "this paper assumes that objects of the same class should all be in a specific region of the image", which is a strong assumption and does not always hold in real-world applications. The author response to this question is not convincing.

Authorsrebuttal2024-08-14

Thanks for the review

Dear Reviewer ra2D, bEdU and Dw1G, Thank you for your insightful comments and constructive suggestions during the review process. Your expertise and attention to detail greatly improve the quality of our manuscript. We are confident that our final submission will address all of the issues you raised and reflect the improvements discussed. Once again, we sincerely thank you for your time and effort and for providing us with the opportunity to improve our work. Best regards, Paper3694 Authors

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC