Segment Everything Everywhere All at Once

In this work, we present SEEM, a promptable and interactive model for segmenting everything everywhere all at once in an image, as shown in Fig.1. In SEEM, we propose a novel decoding mechanism that enables diverse prompting for all types of segmentation tasks, aiming at a universal segmentation interface that behaves like large language models (LLMs). More specifically, SEEM is designed with four desiderata: i) Versatility. We introduce a new visual prompt to unify different spatial queries including points, boxes, scribbles and masks, which can further generalize to a different referring image; ii) Compositionality. We learn a joint visual-semantic space between text and visual prompts, which facilitates the dynamic composition of two prompt types required for various segmentation tasks; iii) Interactivity. We further incorporate learnable memory prompts into the decoder to retain segmentation history through mask-guided cross-attention from decoder to image features; and iv) Semantic-awareness. We use a text encoder to encode text queries and mask labels into the same semantic space for open-vocabulary segmentation. We conduct a comprehensive empirical study to validate the effectiveness of SEEM across diverse segmentation tasks. Notably, our single SEEM model achieves competitive performance across interactive segmentation, generic segmentation, referring segmentation, and video object segmentation on 9 datasets with minimum 1/100 supervision. Furthermore, SEEM showcases a remarkable capacity for generalization to novel prompts or their combinations, rendering it a readily universal image segmentation interface.

Paper

References (90)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer gMmy6/10 · confidence 5/52023-07-01

Summary

The paper proposes SEEM, a method that unifies several segmentation tasks. The network takes text and/or different kinds of visual prompts as inputs. The model then outputs the corresponding segmentation masks for the referred objects. In particular, SEEM trains a unified prompt embedding space. A set of learned queries then alternately attends to image features and the prompt embeddings. Finally, the output embeddings are matched to the prompts and decoded into masks.

Strengths

- Unifying several segmentation tasks. - Impressive results on most tasks. - Similar or better than SAM in many cases, despite less training data. - Reasonably well written.

Weaknesses

1. Some parts of the method lack motivation and details. In particular, the attention mask used in the prompt attention is not well motivated. It is also not well explained that (and how) the output embeddings are matched to the prompts. 2. The paper is heavily derived from X decoder. The magnitude of the contributions is not clear. The authors do not discuss the relation to X decoder. In fact it is hardly mentioned in the paper, yet the method is completely based on it. Seems like the main new addition here is the visual prompt encoder. 3. The results for VOS are quite poor. This indicates that the method is not as robust to object appearance variations, which is crucial in many tasks. 4. The authors should report training and inference time and memory requirements. I am leaning positive for the paper. However, the authors need to address my concerns above.

Questions

See weaknesses. Moreover: Is the method robust to the case when the prompted text does not match any object in the image?

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

Limitations are not discussed in the main paper, but should be mentioned.

Reviewer x17t8/10 · confidence 4/52023-07-06

Summary

This paper introduces SEEM, an innovative and interactive model for segmentation. SEEM stands out with its versatility in handling various prompts, such as points, boxes, scribbles, masks, and texts. The model's design is elegantly simple yet highly effective. Additionally, SEEM incorporates memory prompts to preserve segmentation history, enhancing its interactivity and overall performance.

Strengths

1. The paper is well-presented, showcasing clear and concise writing. The authors effectively communicate their ideas. 2. The proposed idea is both innovative and highly effective. The designed model, despite its simplicity, showcases remarkable capabilities when dealing with a wide range of prompts. Its versatility and adaptability make it be applicable with more tasks. 3. The experimental results are solid. Through a combination of qualitative and quantitative analyses, the authors demonstrate the effectiveness of their idea.

Weaknesses

1. Do the authors observe whether training on all types of prompts performing better than on one type or a subset types of prompts? or which types of prompt performs best and which one still lacks of improvement. It would be better to perform this ablation because it can give more insights to the society. 2. since this is a very general framework supporting many different types of prompts, and it can be more easily extended with more training data. I would also like to see its scalability with larger amount of supervised training data if there is sufficient computing resources.

Questions

Do the authors plan to train on larger amount of dataset? (such as the one trained on SAM or other rich datasets with the semantic labels). It would be very interesting to see how SEEM scale up with large amount of supervised or weakly supervised labels.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and 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

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

not mentioned

Authorsrebuttal2023-08-14

Hi reviewer x17t, thanks for your reviews again! And hope the rebuttal material solve your confusion. It would be really appreciated if you are willing to give any additional comment, and feel free to drop any question if you have any confusion on the rebuttal material : )

Reviewer x17t2023-08-14

Thanks for your effort during the rebuttal. The experimental results are very satisfying and promising. I will slightly improve my rating after the rebuttal. If the authors have time, It would be better to scale up the the model with larger dataset and share the results with the whole community, although it is not necessary to do it during the rebuttal due to the limited time.

Authorsrebuttal2023-08-15

Thanks for your encouraging comment! We will try our best scale up SEEM to larger dataset, it is both beneficial to the community and our work itself : )

Reviewer R6GZ5/10 · confidence 5/52023-07-07

Summary

[Task] In this work, authors introduc SEEM, a promptable and interactive model designed for comprehensive image segmentation. SEEM aims to segment all objects in an image simultaneously, addressing various segmentation tasks. The key contribution of SEEM is its novel decoding mechanism, which allows for diverse prompting and behavior similar to large language models (LLMs). [Method] SEEM is built with four key objectives: versatility, compositionality, interactivity, and semantic-awareness. The model incorporates a visual prompt that unifies different spatial queries such as points, boxes, scribbles, and masks, enabling generalization to different referring images. It also learns a joint visual-semantic space for dynamic prompt composition in different segmentation tasks. Additionally, SEEM incorporate learnable memory prompts in the decoder to retain segmentation history and utilizes mask-guided cross-attention from the decoder to image features. Moreover, a text encoder is employed to encode text queries and mask labels into the same semantic space, facilitating open-vocabulary segmentation. [Experiments] Empirical studies validate the effectiveness of SEEM across diverse segmentation tasks. The model achieves competitive performance in interactive segmentation, semantic segmentation, referring segmentation, and video object segmentation across nine datasets, requiring only minimal 1/100 supervision. SEEM demonstrates remarkable generalization capability to novel prompts or their combinations, positioning it as a versatile and universal image segmentation interface.

Strengths

[highlight the trend towards more flexible segmentation models] In this work, the authors address the need for a universal segmentation interface capable of handling different types of human prompts and addressing various segmentation tasks. They highlight the trend towards more flexible segmentation models, including open-vocabulary segmentation, referring segmentation, and interactive segmentation. Inspired by the success of Large Language Models (LLMs) as universal interaction interfaces for language tasks, the authors propose SEEM, a promptable and interactive model for segmenting everything everywhere all at once in an image. [technical contribution] The proposed approach, SEEM, introduces a novel prompting scheme in the mask decoder with four key properties: versatility, compositionality, interactivity, and semantic-awareness. By encoding different types of prompts (points, masks, text, boxes, and referred regions) into a joint visual-semantic space, SEEM achieves strong compositionality and can handle any combination of input prompts. Memory prompts are introduced to retain previous segmentation information and enable interactivity. The model also provides open-set semantic labels for output segmentation. [tasks and results] The model is trained on diverse segmentation tasks to learn how to handle different prompts, align visual and text prompts, and promote their synergy through cross-attention. The single pre-trained SEEM model achieves competitive performance across all segmentation tasks, leveraging the joint visual-semantic space for prompt combination and zero-shot adaptation. In addition to its strong generalization capability, SEEM is efficient for interactive segmentation compared to other methods. Overall, SEEM offers a segmentation interface with a single pre-trained model that can handle all types of prompts, segment every object with semantics, and cover every pixel in the image.

Weaknesses

[Model performance] The model performance of SEEM on certain benchmarks, including ADE20K for open-vocabulary panoptic segmentation and DAVIS for video instance segmentation, is often lower, and in some cases, significantly lower than the baselines X-Decoder, ODISE, and UNINEXT. It is worth noting that SEEM's framework follows a similar approach to X-Decoder. While the concept of a universal and interactive segmentation interface is intriguing, it appears that the model performance on open-vocabulary benchmarks is compromised in SEEM. It would be beneficial for the authors to provide an explanation for this discrepancy. [Unfair comparisons] In Table 2, there are some comparisons that seem unfair for two reasons. Firstly, the evaluation of SEEM on COCO, which provides instance-level annotations, aligns better with SEEM's training pipeline that includes instance-level annotations. However, this evaluation method puts SAM at a disadvantage since it is trained to respect hierarchical segmentation, including whole-part-subpart relationships. Secondly, SEEM is already trained on in-domain data from COCO's training set, while SAM is pretrained on out-of-domain data. This discrepancy in training data can impact the performance comparison between the two models. To ensure fair comparisons, it would be beneficial to evaluate both models using evaluation protocols that align with their respective training pipelines and take into account the domain of the training data.

Questions

[Handling dataset imbalance] Since SEEM is trained on multiple datasets, there might be an imbalance in the number of samples per dataset. I'm curious to know if adjusting the frequency of training samples from each dataset has any impact on the model's performance. Can the authors shed some light on whether they have explored techniques such as data augmentation or sample weighting to address this issue? Additionally, it would be interesting to understand if adjusting the training sample frequency has any effect on mitigating the impact of dataset imbalance and improving the overall model performance.

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

Yes, the authors adequately addressed the limitations

Authorsrebuttal2023-08-14

Hi reviewer R6GZ, thanks for your reviews again! And hope the rebuttal material solve your confusion. It would be really appreciated if you are willing to give any additional comment, and feel free to drop any question if you have any confusion on the rebuttal material : )

Reviewer cQYZ7/10 · confidence 4/52023-07-29

Summary

This paper proposes a universal segmentation model, SEEM, for all segmentation tasks. A visual sampler module unifies different kinds of human inputs and they are encoded into a joint visual-semantic space together with image and text so that the SEEM model can learn semantic labels for masks. The proposed SEEM model achieves competitive performance across interactive segmentation, generic segmentation, referring segmentation, and video object segmentation. Besides, SEEM shows a strong ability to generalization to novel prompts or their combinations.

Strengths

1. The proposed model has superior performance and has achieved competitive results across different types of segmentation tasks. 2. The proposed model can accept a variety of visual prompts input types and has a strong generalization ability, which makes it possible to be applied to more scenarios. 3. Different from the previous segmentation model, the proposed model can learn the semantic labels corresponding to the segmentation masks, thus enabling open vocabulary segmentation.

Weaknesses

1. In this paper, when comparing the performance of the proposed model with the previous models, only the scales of training data are compared, without the comparison of parameters and calculation. This may cause some doubts about the comparison fairness. 2. The analysis of some results is insufficient. In the ablation study section, the performance changes caused by the change of variables are not consistent in different segmentation tasks. Only the changing trend of the results is given without analyzing the potential reasons.

Questions

1. Although SEEM can learn the semantic labels of masks, it is limited to the categories of COCO. Is there any way to expand the scope of semantic labels without introducing a large amount of extra data? 2. Can you provide a comparison with other universal segmentation models in terms of the number of model parameters and calculations? If this comparison is not applicable, what are the reasons?

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

3 good

Contribution

3 good

Limitations

Limitations are not explicitly discussed in this paper. The scale of training data is limited, and it does not support partial segmentation as in SAM (Segmenting Anything Model).

Authorsrebuttal2023-08-14

Hi reviewer cQYZ, thanks for your reviews again! And hope the rebuttal material solve your confusion. It would be really appreciated if you are willing to give any additional comment, and feel free to drop any question if you have any confusion on the rebuttal material : )

Reviewer cQYZ2023-08-15

Thanks to the authors for the response. My concerns have been addressed and I will keep my positive rating.

Authorsrebuttal2023-08-20

Hi reviewer cQYZ, Thanks so much for your review and positive comments! It would be really appreciated if you could **confirm your final recommendation rating before the final deadline on Aug 21st 1 pm EDT** : ) Thanks a lot!

Authorsrebuttal2023-08-10

Clarification of Q4 and Q5

**Q4: Training and evaluation time and memory cost.** | | Resolution | #Param | FPS/img | Training Memory | Inference Memory | |----------|------------|--------|---------|-----------------|------------------| | SEEM (T) | 1024 | 101.1M | 9.0 | 4804M | 3124M | | SEEM (B) | 1024 | 140.7M | 6.6 | 5556M | 3184M | | SEEM (L) | 1024 | 415.3M | 3.7 | 6588M | 4612M | We compute the FPS on 1 V100 gpu with a single image with resolution 1024 x 1024. Both the training memory and inference memory are referring to batch size equal to 1. Specifically for model training, we fix the visual encoder and language encoder, only the SEEM-Decoder part is learnable. **Q5: Generalization capability on text not included in the image.** Yes, our method is robust to the case when the prompted text does not match any object in the image. Our matching is implemented by computing the similarities between the prompt embedding and object features. We are able to use a threshold to filter those object features with very low similarities. The task we are evaluating is referring segmentation. To further demonstrate this aspect quantitatively, we run our model SEEM-Large on COCO val 2017 of 5000 images. For each of the 5000 images, we sample a positive text prompt and a negative text prompt, and perform the text-based referring segmentation. For each prompt, we calculate the scores over $K$ object tokens using the following calculations: $v = v / norm(v) \in \mathcal{R}^{K \times d}; t = t / norm(t) \in \mathcal{R}^{1 \times d}$ $score = t \cdot v^t / scale \in \mathcal{R}^{1 \times K}$ $score_{max} = \max(score, dim=1)$ where $v$ is the $K$ object query embeddings, $t$ is the textual embedding, and $d$ the feature dimension. We use the maximal score $score_{max}$ as the matching score. Based on the above calculations, we obtain 5000 matching scores for positive prompts and negative ones, respectively. Then we compute the mean and standard deviation as shown below: | Text Prompt Type | Mean | Std | | -------- | -------- | -------- | | Positive | 17.88 | 1.43 | | Negative | 7.36 | 3.00 | Clearly, there is a huge margin between the matching scores for positive and negative text prompts, which justify our above statement.

Reviewer gMmy2023-08-13

I thank the authors for their responses. **Q1:** I understood these things after going back and forth over the manuscript a few times. My point was that, I think the authors could explain these things in a better and more clear way in the paper. This would make the method more easy to comprehend. **Q2:** It is quite clear that the authors directly builds on top of the X decoder. Also if looking at the similarity of presentation. This is not itself an issue. The issue is that it seems that the authors want to hide this. I think this is also harms the readability of the paper. In fact, the proposed method by the authors made much more sense to me after I head the X decoder paper in detail. I therefore would strongly suggest the authors to add a "Preliminaries" section or similar, first briefly describing the X decoder idea. I think this would greatly benefit the presentation of the method, as well as the motivation of the contributions. I think that the contributions wrt X decoder is sufficient. **Q5:** I thank the authors for this experiment. However, it would be more interesting if the authors could provide an analysis of the percentage of frames where a negative prompt lead to a detected mask (for some good global value of the threshold). Or to provide some precision-recall analysis. In addition to text prompt, it would also be interesting if the authors could perform a similar evaluation for image prompts (VOS or visual referring segmentation).

Authorsrebuttal2023-08-14

Thanks for the prompt reply : ) Below are some quick responses, we are working on the experiment and will update shortly for Q5. **Q1**: Thanks for your efforts on going through the paper iteratively, and this is a very valuable suggestion. We will definitely include a more clear explanation on the attention mask and output embeddings in the future versions. To make sure our work is reproducible for the community, we will release the training/evaluation code and models. **Q2**: We sincerely apologize for making the paper seem to be presented in this way and thank you for recognizing the contribution of our work with respect to X-Decoder. It is a great advice to add a "preliminary" section about X-Decoder before presenting our SEEM. We will follow the suggestion and carefully integrate this into the next version. In addition, we would like to clarify that it is an honest mistake instead of we are pretending not building SEEM upon X-Decoder. Again we appreciate that you point this out and give us a chance to have a better presentation for future versions. **Q5**: Great suggestion! we will provide the results shortly.

Authorsrebuttal2023-08-15

**Q5**: We are providing the accuracy of the threshold that is able to successfully separate the positive and negative example in cross-validation format on both text prompt and visual prompt. 5-Fold cross validation results for text referring segmentation: Validation Fold | Fold-1 | Fold-2 | Fold-3 | Fold-4 | Fold-5 | | -------- | -------- | -------- | -------- | -------- | -------- | EER Threshold | 13.57 | 13.77 | 13.74 | 13.67 | 13.67 | Accuracy | 98.43 | 98.84 | 98.94 | 98.23 | 98.54 | 5-Fold cross validation results for image referring segmentation: Validation Fold | Fold-1 | Fold-2 | Fold-3 | Fold-4 | Fold-5 | | -------- | -------- | -------- | -------- | -------- | -------- | EER Threshold | -0.14 | -0.19 | -0.19 | -0.19 | -0.19 | Accuracy | 74.17 | 70.83 | 72.50 | 73.33 | 71.67 | * We are using DAVIS17 for evaluating the image referring segmentation, the positive object are sampled from the same video of the same object id, and the negative object are randomly sampled from another video. EER - Equal Error Rate Accuracy - The accuracy of positive and negative examples given the EER Threshold The result clearly shows that our model is able to separate the positive and negative example given a threshold.

Authorsrebuttal2023-08-20

Hi Reviewer gMmy01, As suggested, we have provided solution for all the questions you raised including the newly required experiments on "the percentage of frames where a negative prompt led to a detected mask." Since the due for the final recommendation is close, it would be really appreciated that **you could confirm whether the questions are resolved and make your final recommendation**. Thanks!

Reviewer gMmy2023-08-22

I thank the authors for the responses. I am happy with the response and results and will increase my rating. I think that the authors should integrate the above feedback and experiments to the main paper and/or supplementary material.

Authorsrebuttal2023-08-22

Thanks so much! We will make the suggested adjustment on the incoming updates!

Reviewer R6GZ2023-08-15

Thank you for your comprehensive response addressing the questions raised during the rebuttal period. I would also like to extend my gratitude to the authors for clarifying some of the confusion I had regarding this paper. Having reviewed the rebuttal, I find that the majority of my questions have been effectively resolved. Consequently, *I am inclined to recommend the acceptance of this paper.* While I acknowledge the significant interest inherent in many of the findings presented in this work, I still have reservations regarding the performance gap between X-Decoder and SEEM, particularly given that SEEM is built upon X-Decoder. Echoing questions raised by other reviewers, I propose that training the model on a large-scale dataset and reporting the results would be nice. I am intrigued to learn whether increased data for training could yield further improvements in model performance.

Authorsrebuttal2023-08-15

Hi Reviewer R6GZ, thanks so much for your encouraging comments! We agree with your comment that while building upon X-Decoder, SEEM has its own advantages (composition capability, emerging property and etc. ) and disadvantages (overall open vocabulary segmentation performance). However, we emphasis that the disadvantages are coming from the task focusing. We have done some pilot study on scaling up segmentation dataset, that has shown generalization ability of our approach (improved NoC, 1-IoU). Thanks again for the reviewer's reply : )

Authorsrebuttal2023-08-20

Hi reviewer R6GZ, Thanks so much for your review and positive comments! We appreciated that you are considering "inclined to recommend the acceptance of this paper" during the discussion period. The current rating is still "Borderline accept", it would be really appreciated if you could **make the final recommendation before the deadline on Aug 21st 1 pm EDT** : ) Thanks for your patience again!

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC