Towards Open-Vocabulary Semantic Segmentation Without Semantic Labels

Large-scale vision-language models like CLIP have demonstrated impressive open-vocabulary capabilities for image-level tasks, excelling in recognizing what objects are present. However, they struggle with pixel-level recognition tasks like semantic segmentation, which additionally require understanding where the objects are located. In this work, we propose a novel method, PixelCLIP, to adapt the CLIP image encoder for pixel-level understanding by guiding the model on where, which is achieved using unlabeled images and masks generated from vision foundation models such as SAM and DINO. To address the challenges of leveraging masks without semantic labels, we devise an online clustering algorithm using learnable class names to acquire general semantic concepts. PixelCLIP shows significant performance improvements over CLIP and competitive results compared to caption-supervised methods in open-vocabulary semantic segmentation. Project page is available at https://cvlab-kaist.github.io/PixelCLIP

Paper

References (67)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 83hz5/10 · confidence 4/52024-07-06

Summary

This paper present MCLIP, which aims to finetune CLIP using image and mask data without semantic labels. The goal is to make its open-vocabulary recognition ability to be adapted to position-sensitive semantic segmentation tasks. MCLIP chooses to use SAM or feature clusters of DINO to obtain masks, which are class-agnostic but sometimes too small to have consistent semantics for finetuning. In addition, in order to achieve a trade-off between training stability and avoiding catastrophic forgetting, clustering uses the EMA version of finetuned image encoder.

Strengths

- The idea of this paper is interesting. Solving open-vocabulary semantic segmentation without semantic labels is an interesting topic. - The experimental results show that the preformance of the proposed approach is good. - Ablation studies and analysis are provided to help better understand of the role of each component and hyper-parameter.

Weaknesses

- What role does the text encoder play in the clustering process? The key to the success of clustering lies in the existence of unlabeled data in a certain grouping relationship in the space. The mask representations f_M can be considered to meet such conditions. If k learnable object prompts are initialized randomly, then k text representations f_C are randomly scattered in the text space initially, unable to provide the information needed for clustering. If k learnable object prompts are initialized by predefined categories, then this should violate the experimental setup without semantic labels. To illustrate this problem, the authors need to provide an explanation and add a comparative experiment: What results can be obtained just by clustering with f_M? - Why use clustering technology to solve the problem of too fine initial mask granularity? Firstly, previous work (GroupViT, CVPR'22 and SegCLIP, ICML'23) has used a grouping block designed based on the idea of clustering in the process of fine-tuning CLIP. In contrast, the clustering used in MCLIP provides better masks initially than the image patches used in previous work and the clustering process of MCLIP is parameter-free. It may be necessary to conduct ablation experiments to prove which differences bring positive results. But overall, technically, the use of clustering may be not novel. - In order to obtain class-agnostic masks, there is a class of experimental settings called open set semantic segmentation (O3S, arXiv preprint arXiv:2307.02003, 2023) . Can the ”SAM initialization + clustering” scheme proposed by MCLIP obtain competitive performance in the O3S experimental setting? - Table 1 shows that using MCLIP’s fine-tuned CLIP instead of the Frozen CLIP in the previous method improves the performance of the previous method’s zero-shot mask classification. Taking FC-CLIP as an example, MCLIP improves its zero-shot mask classification performance on ADE20K by 2.7. But compared to the final performance of FC-CLIP with two branches integrated, it is still 4.0 lower. So, in the case of keeping the complete inference strategy of the previous method, how much improvement can be achieved by using MCLIP’s fine-tuned CLIP instead of the Frozen CLIP? - In the ablation experiment in Table 3a, does ’w/o Semantic Clustering’ mean fine-tuning with SAM-initialized masks directly without clustering? If so, why is the performance so poor? Without using clustering, it is equivalent to the number of cluster k is greater than the number of masks initialized by SAM on all images. However, the conclusion obtained from table 3b is that with the increase of k, the performance tends to saturation, rather than decreasing to the level of ”w/o Semantic Clustering”.

Questions

I think the authors need to further explain the novelty of this paper and give more analysis on the experimental results as listed in the weaknesses part.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

Limitations are provided. It seems that there is no potential negative societal impact.

Reviewer kepm6/10 · confidence 4/52024-07-08

Summary

The authors introduce a novel unsupervised formulation of open vocabulary semantic segmentation, adapting a pre-trained vision-language model (CLIP) to the task via distillation from vision-only segmentation models (i.e., SAM, DINO). To use the language encoder without constraining fine-tuning to the set of classes of the dataset, they propose to apply online clustering at the dataset level and to learn class-specific feature prototypes. The method is compared against state-of-the-art approaches trained with and without supervision, and the results validate the approach as a practical direction for training open vocabulary models for semantic segmentation.

Strengths

- The work is well-presented and curated, and the motivation is clear and sound. - The approach finetunes models pre-trained for a different purpose, and employs supervision from other pre-trained models, effectively reusing knowledge efficiently. Distillation from pre-trained models also removes the need for data supervision, potentially permitting scale training for open vocabulary semantic segmentation to billions of samples. - The method is tested against various baseline and on different benchmark datasets. - The model components are evaluated independently in the ablation study, helping to uncover the individual contribution to the final picture.

Weaknesses

- While I understand the rationale behind performing clustering at pixel level on the entire dataset, I am not sure of the scalability of the approach. This probably explains why the authors only use 5% of the SA-1B dataset. It would be helpful to quantify the costs of performing the online clustering at training time. - Since the method could potentially suffer scalability issues, it would be interesting to understand performance when trained unsupervised in-distribution, reporting both in- and out-of-distribution performance, e.g., MCLIP trained on VOC and tested on Context. - The reasoning behind the selection process for the prompt (i.e., "a photo of a {} in the scene") is unclear. This is a minor issue, but the selection should probably be justified. While I would expect the model to learn to ignore the prefixes and suffixes, it would be interesting to understand how the model performs with other prompts. - Using CLIP as a baseline for the ablation studies may be misleading due to the mismatch between its application and open vocabulary semantic segmentation

Questions

- What are the costs of performing online clustering at the dataset level during training? - What is the performance of MCLIP when trained on the benchmark datasets and tested in- and out-of-distribution? How does it compare against some baseline methods? - What is the model performance when changing the template at inference time? What is the performance without the template? What about training without the template? Does performance improve with prompt ensembling (i.e., similar to what CLIP does with N prompts that average to get better class centroids)?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

- A potential limitation is the one I reported above, i.e., scalability issues due to the global online clustering procedure.

Reviewer rHvB5/10 · confidence 4/52024-07-14

Summary

This paper proposes to learn an open-vocabulary semantic segmentation model with only unlabeled images and pretrained foundation models, such as SAM and DINO. The intuition is that CLIP model already knows what is in the image, so we only need to teach CLIP where the object is. It first uses pretrained DINO to generate pseudo masks and then exploits a online clustering method to group the part segments into valid object masks. It also proposes learnable class embeddings to solve the problem of lacking ground-truth text labels. Compared with baselines, the proposed method achieves decent improvements.

Strengths

+ This paper proposes a solution to train an open-vocabulary model only with unlabeled images, i.e., without masks nor captions. + Three technical contributions to make the solution happen: (1). use DINO to generate pseudo masks (2). group part masks into objects (3). user learnable embeddings to substitute text captions.

Weaknesses

- Compared with other methods [ 50 , 30 , 43 , 51 ] leveraging image caption as supervision, this paper actually uses stronger DINO generated pseudo masks to train the segmentation model. Furthermore, it even uses the SAM masks during experiments (In Table 2). Regrading this, I think this method includes strong segmentation prior into the training, making the comparison unfair. - Compared with methods using similar VFM, such as SAM-CLIP, the proposed method performs much worse. I understand it is not an apple-to-apple comparison due to the differences in training data. More specially, this paper doesn't use captions. However, caption is also very easy to get with pretrained caption models. If simply adding caption could boost the performance so much, why should we stick to a setting without caption?

Questions

- Why does Table 1 appear before Table 2? Tabel 1 looks more like an ablation study to me.

Rating

5

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

See weakness

Reviewer Dx6v6/10 · confidence 4/52024-07-15

Summary

This paper proposes to enhance the semantic segmentation performance of the pretrained CLIP model using unlabeled images and pseudo segmentation masks generated with vision foundation models such as SAM and DINO. Specially, the pseudo masks are acquired via a online feature clustering algorithm. Experiments on standard benchmarks demonstrate the superior performance over CLIP and competitive results to existing open-vocabulary semantic segmentation methods.

Strengths

1. It is a neat idea to leverage unlabeled masks as supervision generated from foundation models (e.g. SAM, DINO) for the open-vocabulary semantic segmentation task. 2. The paper is well-written and easy to follow. 3. The ablation study demonstrates the effectiveness of some design choices such as momentum update of the image encoder, online cluster assignment, and learning class prompts. Performance on the standard benchmarks also show the competitive results in comparison with existing baseline methods.

Weaknesses

1. How is the proposed method in comparison with MaskCLIP+? Figure 4 shows the visual comparisons between the proposed method and MaskCLIP, as the Sec. 4.2 mentioned “For evaluating CLIP [42, 10], we apply MaskCLIP [61] to extract CLIP image features”. It would be interesting to compare the proposed method with MaskCLIP+ which distills MaskCLIP to train more advanced segmentation model. 2. Some related works should be discussed and compared to the proposed method, for example, Exploring Open-Vocabulary Semantic Segmentation from CLIP Vision Encoder Distillation Only. ICCV 2023.

Questions

See the weakness section.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes

Reviewer rHvB2024-08-07

I've raised my score from 4 to 5

Thanks for the rebuttal! I've raised my score from 4 to 5.

Reviewer kepm2024-08-08

I thank the authors for their reply, which answers my doubts and questions. I appreciate that they executed all the experiments I proposed. Scaling the dataset size seems to (unsurprisingly) improve model performance. Also, a better template for the input query or a template ensemble has a positive impact. For the first, I believe the work would benefit from some experiments showing how performance increases with an increase in the number of train samples. I expect the performance to improve until the compute-optimal point is reached [1]. Overall, I am glad my suggestions improved the method's performance, however, this also gives the impression the work was not carried out very rigorously. In any case, I am satisfied with the authors' rebuttal and confirm my initial positive rating. [1] Hoffmann, Jordan, et al. "Training compute-optimal large language models." arXiv (2022).

Reviewer 83hz2024-08-12

Final rating

I appreciate the responses from the reviewers. Though most of my concerns have been solved, I still think the novelty of this paper cannot reach the level of an accept. So, I tend to keep my original rating unchanged.

Reviewer Dx6v2024-08-12

Thanks for the detailed response! I'll raise my initial rate to weak accept.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC