Rewrite Caption Semantics: Bridging Semantic Gaps for Language-Supervised Semantic Segmentation

Vision-Language Pre-training has demonstrated its remarkable zero-shot recognition ability and potential to learn generalizable visual representations from language supervision. Taking a step ahead, language-supervised semantic segmentation enables spatial localization of textual inputs by learning pixel grouping solely from image-text pairs. Nevertheless, the state-of-the-art suffers from clear semantic gaps between visual and textual modality: plenty of visual concepts appeared in images are missing in their paired captions. Such semantic misalignment circulates in pre-training, leading to inferior zero-shot performance in dense predictions due to insufficient visual concepts captured in textual representations. To close such semantic gap, we propose Concept Curation (CoCu), a pipeline that leverages CLIP to compensate for the missing semantics. For each image-text pair, we establish a concept archive that maintains potential visually-matched concepts with our proposed vision-driven expansion and text-to-vision-guided ranking. Relevant concepts can thus be identified via cluster-guided sampling and fed into pre-training, thereby bridging the gap between visual and textual semantics. Extensive experiments over a broad suite of 8 segmentation benchmarks show that CoCu achieves superb zero-shot transfer performance and greatly boosts language-supervised segmentation baseline by a large margin, suggesting the value of bridging semantic gap in pre-training data.

Paper

Similar papers

Peer review

Reviewer wvpf6/10 · confidence 3/52023-07-04

Summary

This paper first points out the semantic gap problem of the existing language-supervised semantic segmentation method. It shows that not all visual elements are included in the corresponding language annotations. Then the paper proposes Concept Curation (CoCu), which includes Vision-driven Expansion, Text-to-Vision-Guided Ranking, and Cluster-guided Sampling strategies to solve the semantic gap problem. The experiments show the effectiveness of the CoCu.

Strengths

1. The semantic gap problem sounds very reasonable and significant for language-supervised semantic segmentation. 2. The proposed method CoCu could alleviate the semantic gap problem to some extent. 3. The experiments and ablation study are comprehensive.

Weaknesses

1. The authors should summarize the Sec. 2.3 at the beginning or the end of this section. For example, in general, Sec. 2.3 provides a method to find better class candidates for the loss in Sec. 2.1. Otherwise it confuses the readers about the ultimate goal of Sec. 2.3. 2. The overall pipeline of CoCu is a kind of dataset pre-processing method, and it is complicated. Is CoCu done online or offline? If online, how much time does it cost to find the final class candidates?

Questions

None

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

3 good

Contribution

3 good

Limitations

None

Reviewer iruH6/10 · confidence 4/52023-07-04

Summary

This paper proposed a novel data curation/argumentation process, named Concept Curation (CoCu), for language-supervised semantic segmentation. In the setting of language-supervised semantic segmentation (e.g. GroupViT), the network is trained with image-text contrastive loss on large-scale image-text pairs. Authors identified several issues in the vanilla data distribution of the original contrastive learning, e.g. semantic gap, and semantic bias. The proposed CoCu mitigates these issues and improves over prior work GroupViT under a controlled setting, showing a faster convergence rate and high accuracy.

Strengths

1. In the quantitative evaluation, authors re-implemented GroupViT and compare CoCu and GroupViT under the controlled experiment setting, i.e. 1024 batch size. The authors also report the accuracies on additional evaluation datasets, IN50, IN300, Cityscapes, etc. CoCu outperforms GroupViT on all the datasets by a margin. Although CoCu doesn't achieve state-of-the-art results on some datasets, it doesn't degrade the effectiveness of the method. 2. The proposed method is well-motivated. The web-crawled image-text pairs are indeed quite noisy. And GroupViT is also known to be bad at segmenting background classes like grass and sky. Moreover, from the visualization in Figure 3(b), the CoCu shows that could focus on background grasses instead of the foreground fox. 3. In the Table 2 ablation study, when only trained on CC3M, CoCu improves over GroupViT by a large margin on Pascal VOC. It is a very interesting result and justifies that CoCu speeds up the convergence.

Weaknesses

1. Insufficient qualitative comparison. Since one major claim of CoCu is that, compared with vanilla contrastive loss used in GroupViT, the proposed dataset curation mitigates semantic cap and semantic bias issues. So besides quantitative evaluation metric mIoU, more visualizations compared with GroupViT are expected. I would suggest authors add more visualizations in the supplementary materials if there is no space left in the main submission. 2. In the ablation table 3, the authors studied the effects of different components of CoCu. But the average mIoU may not be an insightful metric. Since different datasets have very different category vocabulary, authors may try to include a detailed ablation study table in the supplementary material and elaborate more on the results.

Questions

1. The boundary between the semantic gap and semantic bias is not that clear. At least from Figure 1, it's hard to tell the major differences between them. Maybe authors could define them more rigorously and provide more examples about them. 2. How long does it take to process the dataset? It would be informative for other readers who want to reproduce or follow up on the curation process. 3. In Table 2, CoCu also showed significant improvement on ImageNet50 and ImageNet300. But for these two datasets, the vocabularies are mainly foreground, it would be very interesting to show why CoCu could improve on these datasets as well.

Rating

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

4 excellent

Presentation

2 fair

Contribution

3 good

Limitations

Yes.

Reviewer griQ6/10 · confidence 4/52023-07-05

Summary

This paper targets learning unsupervised semantic segmentation from image-text pairs. The authors specifically address the issue of training data quality in previous method Group-ViT and propose an approach to enhance the captions with additional visual concepts through an automated pipeline. The paper demonstrates the effectiveness of this data filtering pipeline by evaluating it on official segmentation benchmarks.

Strengths

- The semantic gap problem the authors studied is intresting and meaning full. - The writing is good and the paper is easy to understand.

Weaknesses

The proposed data filtering pipeline relies on the CLIP model for collecting visually similar samples. However, the CLIP model has been trained on a much larger scale, with hundreds of millions of image-text pairs, whereas the training data used in this paper is relatively smaller. Does it work by distilling the CLIP model? This raises concerns about **the effectiveness of the pipeline when scaling up to larger training data sizes**. It would be helpful to explore **alternative self-supervised methods such as DINO, MAE, or Group-ViT itself to replace the CLIP model** in the pipeline and evaluate its performance.

Questions

See weakness.

Rating

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

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

NaN

Reviewer e4vq5/10 · confidence 5/52023-07-06

Summary

Current VLMs suffer from a noticeable semantic gap between visual and textual modalities since many visual concepts present in images are easily missed in their paired captions. This work proposes Concept Curation (CoCu), a pipeline that leverages CLIP to compensate for the missing semantics. For each image-text pair, CoCu establishes a concept archive that maintains potential visually-matched concepts using vision-driven expansion and text-to-vision-guided ranking. This approach enables the identification of relevant concepts through cluster-guided sampling, which are then fed into the pre-training process. As a result, CoCu bridges the gap between visual and textual semantics. Extensive experiments conducted on eight segmentation benchmarks demonstrate that CoCu achieves exceptional zero-shot transfer performance and significantly enhances the language-supervised segmentation baseline by a substantial margin. These results underscore the importance of closing the semantic gap in pre-training data. The code for CoCu will be made available to the research community.

Strengths

1. Overall, the paper is well-written, and the study of this work has a good motivation. 2. The mechanism of CoCu is intuitive and easy to follow. 3. CoCu demonstrates significant empirical results: it outperforms GroupViT (re-implemented baseline) by 4.6% mIoU in average on eight popular semantic segmentation benchmarks. 4. CoCu also yield good qualitative results, with more accurate and smooth masks than its baseline.

Weaknesses

1. Despite the fact that CoCu indeed enriches visual concepts during VL pre-training, the cost seems to be heavy. During training, a CLIP model is employed to perform the key components of CoCu such as vision-driven expansion and text-to-vision guided ranking, which introduces a lot of additional computation. Thus, the comparison to your baselines such as GroupViT might not be strictly fair. A detailed comparison of training time or FLOPS should be presented to support the effectiveness of CoCu. 2. My bigger concern lies in how CoCu relies on the pre-trained CLIP, i.e., if using a weaker CLIP model for CoCu, how will the performance change? If CoCu is not sensitive to the CLIP's performance, you can directly use GroupViT to perform vision-driven expansion and text-to-image-guided ranking, so that no additional parameters will be introduced. As shown in Table 4, GroupViT obtains acceptable zero-shot classification results. However, if CoCu highly relies on a strong CLIP, your contributions might also be challenged.

Questions

1. In table 1, are both CoCu and GroupViT trained for 30 epochs? 2. With the concern of Weakness 1, I am also wondering what if you change the x-axis from #epochs to actual training time in Figure 3? Will CoCu still converge faster than GroupViT?

Rating

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

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

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

As CoCu basically follows the architecture of GroupViT, it has the same limitation that for each image, the number of semantic concepts segmented by CoCu/GroupViT has a maximum of the number of group tokens fed to the vision encoder. Thus, given a high-resolution image with many semantic regions, these methods might fail to generate accurate segmentation masks.

Reviewer wvpf2023-08-14

Response to the rebuttal

Thanks for the answer. I keep my rating.

Area Chair Dbh42023-08-18

Hi Authors, Thanks a lot for your rebuttal. We are urging all the reviewers to respond and will take your input into consideration as we make the final recommendation. Thanks! Best,

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC