One Token to Seg Them All: Language Instructed Reasoning Segmentation in Videos

We introduce VideoLISA, a video-based multimodal large language model designed to tackle the problem of language-instructed reasoning segmentation in videos. Leveraging the reasoning capabilities and world knowledge of large language models, and augmented by the Segment Anything Model, VideoLISA generates temporally consistent segmentation masks in videos based on language instructions. Existing image-based methods, such as LISA, struggle with video tasks due to the additional temporal dimension, which requires temporal dynamic understanding and consistent segmentation across frames. VideoLISA addresses these challenges by integrating a Sparse Dense Sampling strategy into the video-LLM, which balances temporal context and spatial detail within computational constraints. Additionally, we propose a One-Token-Seg-All approach using a specially designedtoken, enabling the model to segment and track objects across multiple frames. Extensive evaluations on diverse benchmarks, including our newly introduced ReasonVOS benchmark, demonstrate VideoLISA's superior performance in video object segmentation tasks involving complex reasoning, temporal understanding, and object tracking. While optimized for videos, VideoLISA also shows promising generalization to image segmentation, revealing its potential as a unified foundation model for language-instructed object segmentation. Code and model will be available at: https://github.com/showlab/VideoLISA.

Paper

Similar papers

Peer review

Reviewer oRTn5/10 · confidence 4/52024-06-24

Summary

This work introduce a Sparse Dense Sampling strategy and a One-Token-Seg-All approach to enhance the temporal ability of LISA. For the Sparse Dense Sampling strategy, the work preserves dense tokens of some frames and extract sparse tokens of interleaved frames. For the One-Token-Seg-All approach, the work apply the feature of one embedding can represent objects across multiple frames. Additionally, this work contributes a reason VOS benchmark.

Strengths

This work proposes a benchmark for reasoning video segmentation, including 458 video-instruction samples.

Weaknesses

1. This work utilizes two visual encoder. It's redundant, and will influence the speed of the model. 2. The model fails to segment multiple objects at the same time.

Questions

1. The results when replace the SAM with other kernel-based segmenting and tracking models. 2. Compared with QFormer [1] and QFormer-like archectures [2], what are advantages of the Sparse Dense Sampling strategy? [1] Li J, Li D, Savarese S, et al. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models[C]//International conference on machine learning. PMLR, 2023: 19730-19742. [2] Li Y, Wang C, Jia J. LLaMA-VID: An image is worth 2 tokens in large language models[J]. arXiv preprint arXiv:2311.17043, 2023.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

1. This work utilizes two visual encoder. It's redundant, and will influence the speed of the model. 2. The model fails to segment multiple objects at the same time.

Reviewer vGHj5/10 · confidence 5/52024-07-08

Summary

This work proposes VideoLISA, a video-based multimodal large language model that addresses the challenges of language-instructed reasoning segmentation in videos, leveraging various strategies to enhance temporal understanding and consistent object tracking, and showing promising generalization capabilities. The proposed Sparse Dense Sampling strategy is able to reduce computational costs, and a specially designed TRK token can segment and track the object across frames. They also establish a benchmark, demonstrating VideoLISA's performance.

Strengths

1. VideoLISA is the first video-LLM that democratizes reasoning segmentation to videos, which generates temporally consistent segmentation masks in videos based on language instructions. 2. The experiments evaluate the model performances from multiple segmentation benchmarks which demonstrated the improvements. 3. Many qualitative results and comparisons are provided.

Weaknesses

1. The token "TRK" is only one token designed for the task, so it seems the model cannot generate multiple segmentation mask tracklets. 2. VideoLISA has limited technical contributions in its design, where only the sampling strategy is specially designed for the video domain, and the "TRK" token seems to be a simple extension from the original LISA with a simple adaptation. 3. Pixellm is another reasoning segmentation work that is able to segment multiple targets and support multi-round conversations, which is not compared in the paper.

Questions

1. How to use this single token to generate segmentation masks for multiple targets? 2. I wonder how the model performs on text-generation task, does the model preserve the original ability to perform conversation? 3. The sampling strategy reminds me of the strategy used in LLaMA-VID, which compresses the visual information of each frame to one token. I am curious about the performance of adding a segmentation head to such models, which is capable of both generating texts and producing masks.

Rating

5

Confidence

5

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes, authors addressed the limitations of their work.

Reviewer urmZ5/10 · confidence 3/52024-07-15

Summary

This paper introduces VideoLISA, a multimodal LLM for reasoning segmentation in videos. A Sparse Dense Sampling strategy is proposed to balance the temporal context and spatial detail for video modeling. Extensive results on various benchmarks demonstrate the effectiveness of the proposed method.

Strengths

The paper is well-organized and easy to follow. The introduction of the VideoLISA model and the ReasonVOS benchmark establishes a new paradigm in video segmentation, making it highly inspiring.

Weaknesses

1. The proposed Sparse Dense Sampling Strategy and One-Token-Seg-All framework are both very simple and lacks sufficient innovation. 2. The setups of some ablation studies in the article are not very clear, e.g. whether the n of **n-frame** in Table5 is T-sparse or T-dense is not clarified, how to combine XMem with LISA in Table 6 is also not introduced.

Questions

Please refer to the weakness part.

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

The limitations have been discussed in the paper.

Area Chair xx8c2024-08-12

Help Check the Rebuttal, Make Discussions, and Update the Final Recommendation

Dear Reviewers, Thanks for serving as a reviewer for the NeurIPS. The rebuttal deadline has just passed. The author has provided the rebuttal, could you help check the rebuttal and other fellow reviewers' comments, make necessary discussions, and update your final recommendations as soon as possible? Thank you very much. Best, Area Chair

Reviewer oRTn2024-08-12

Thanks to the author's rebuttal, which has completely solved my doubts; however, as other reviewers mentioned, after training with reasoning segmentation, the model loses the ability of text-generation, which needs to be noted.

Authorsrebuttal2024-08-12

Dear Reviewer oRTn, Thank you very much for your positive feedback! We are glad to hear that our rebuttal has completely resolved your concerns! Regarding the minor limitation you mentioned, as we discussed in our rebuttal, degraded text generation capability is a common problem shared by many models (e.g., LISA). Maintaining the two functions in one model is a meaningful yet independent research problem, involving data curation, training strategy design, etc. We are glad to include this discussion into our revised paper to inspire future works. Yet, we believe this common limitation does not affect our contributions, as our VideoLISA focuses on processing videos, which is orthogonal to the compatibility of text-generation. We thank the reviewer again for the valuable suggestions, which are very helpful in improving our work.

Reviewer vGHj2024-08-12

Thank you for the thorough responses. The responses partially address my concerns. However, there are still a few points that are not convincing. - Capacity of segmenting multiple objects. I understand segmenting multiple instances is much more difficult than segmenting a single object. However, the name of your approach, "one token seg all", emphasizes the capacity to segment multiple objects at once with a single token. - As for the comparison of PixelLM with your approach, can you provide further analysis of the performance gap? Because PixelLM is capable of handling multiple objects while yours isn't. The major difference seems to be that you have the temporal association module while previous methods don't. However, temporal association (via one token) is not novel in previous video segmentation tasks.

Authorsrebuttal2024-08-12

Dear Reviewer vGHj, Thank you very much for your positive feedback and valuable suggestions! We are glad to hear that our rebuttal has addressed some of your concerns. Regarding the two points you mentioned: - We would like to clarify that "One-Token-Seg-All" emphasizes the capacity of using one single token to segment **multiple video frames**, rather than **multiple objects**. We also do not claim the ability to handle multiple objects in our paper. - As we discussed in the rebuttal, single-object referring/reasoning segmentation is a standard and widely adapted setting. Since our main focus is video data, rather than single versus multi-object, we believe this standard setting is suitable and effective to assess our method. - We greatly thank the reviewer for pointing out this potential misleading expression in the name of the approach. We will consider revise the name and add more clarification in the revised paper. - In the comparison between VideoLISA and PixelLM on video object segmentation benchmarks, VideoLISA shows better performance than PixelLM. The performance gap mainly comes from **the model's design on processing videos**. We provide the analysis below and will add this in our revised paper. - PixelLM focuses on multi-object in **images** but has little design for videos. When applying such image-based models on videos, they present two challenges: 1) cannot understand video temporal dynamics, making it unable to handle temporal-related text queries, such as actions, state changes, etc. 2) cannot output temporal-consistent masks. The two issues make image-based models, including LISA and PixelLM, struggle to properly process video data. - Assume that even we equip the image-based model with a "temporal association module", it is still hindered by the first challenge on video understanding. **This argument has been validated by experiments in our paper**. In Table 6 of the paper and Table 9 of the appendix, we tried to equip LISA with a "temporal association module" XMem. It still shows significant worse performance than VideoLISA. - In contrast, our VideoLISA is dedicatedly designed for videos. Rather than a "temporal association module", our VideoLISA actually has **two novel designs for videos**, including a Sparse Dense Sampling strategy and a One-Token-Seg-All approach. The design motivation and rationale have been discussed in the rebuttal. The effectiveness has been validated by extensive experiments in our paper. Corresponding to the two challenges stated above, both of the two modules are essential. The Sparse Dense Sampling makes the model be aware of both spatial and temporal information in videos while the One-Token-Seg-All enables temporal consistent segmentation. - From the functionality perspective, PixelLM addresses image understanding, language reasoning; previous video segmentation models may mainly address image (frame) understanding and video temporal association. **Our VideoLISA is the first model that systematically integrates all these features**, and this is non-trivial to achieve (not simply stacking these existing models together). As mentioned above, the experiment in Table 9 of the appendix also demonstrates that simply stacking these existing models only yields sub-optimal results. We greatly thank the reviewer for the feedback and constructive questions, which are very helpful in improving our work. We sincerely hope this response answers your questions. Any questions/comments are warmly welcomed!

Reviewer vGHj2024-08-13

Thank you for your response. Since most of my concerns have been addressed, I will raise the score to BA. However, in the main text, the discussion and comparison of different image reasoning methods with different temporal association designs (other than XMem) are quite limited. Also, the experiments comparing the dense-sparse sampling strategy with the QFormer structure should be included to address the novelty. What's more, the capability of degraded text generation should be discussed and mentioned (For example, PixelLM can have conversations with users while LISA can not). I hope the final version will be more organized in this regard.

Authorsrebuttal2024-08-13

Dear Reviewer vGHj, We are glad to hear that our response has successfully addressed your concerns! We sincerely thank you for your constructive suggestions. We will follow your suggestions to include more comprehensive comparisons in our revised manuscript, including different image reasoning segmentation methods with different temporal association designs, and Q-Former architecture, and so on. Moreover, we will discuss the problem of text generation capability to inspire more future works. Thank you again for your great efforts to help us improve!

Reviewer urmZ2024-08-13

Response to rebuttal

Thanks for the thorough explanation. The responses address most of my concerns, I will raise my rating to borderline accept.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC