Self-Chained Image-Language Model for Video Localization and Question Answering

Recent studies have shown promising results on utilizing large pre-trained image-language models for video question answering. While these image-language models can efficiently bootstrap the representation learning of video-language models, they typically concatenate uniformly sampled video frames as visual inputs without explicit language-aware, temporal modeling. When only a portion of a video input is relevant to the language query, such uniform frame sampling can often lead to missing important visual cues. Although humans often find a video moment to focus on and rewind the moment to answer questions, training a query-aware video moment localizer often requires expensive annotations and high computational costs. To address this issue, we propose Self-Chained Video Localization-Answering (SeViLA), a novel framework that leverages a single image-language model (BLIP-2) to tackle both temporal keyframe localization and QA on videos. SeViLA framework consists of two modules: Localizer and Answerer, where both are parameter-efficiently fine-tuned from BLIP-2. We propose two ways of chaining these modules for cascaded inference and self-refinement. First, in the forward chain, the Localizer finds multiple language-aware keyframes in a video, which the Answerer uses to predict the answer. Second, in the reverse chain, the Answerer generates keyframe pseudo-labels to refine the Localizer, alleviating the need for expensive video moment localization annotations. Our SeViLA framework outperforms several strong baselines on 5 challenging video QA and event prediction benchmarks, and achieves the state-of-the-art in both fine-tuning (NExT-QA, STAR) and zero-shot (NExT-QA, STAR, How2QA, VLEP) settings. We also analyze the impact of Localizer, comparisons of Localizer with other temporal localization models, pre-training/self-refinement of Localizer, and varying the number of keyframes.

Paper

Similar papers

Peer review

Reviewer mWtS6/10 · confidence 4/52023-06-17

Summary

This work builds a joint model for temporal language grounding and video question answering. The model is built on a state-of-the-art image-language model, BLIP-2, and finetunes it in a parameter-efficient way to derive a localizer and an answerer. The localizer finds language-aware keyframes in a video and the answerer uses these frames to predict the answer. The answerer is also used to generate pseudo-labels to refine the localizer. Experiments on multiple video question answering datasets in both finetuning and zero-shot settings show that the proposed method achieves state-of-the-art results and improves slightly over the BLIP-2 baseline.

Strengths

- Leveraging image-language models for video-language tasks is an important direction, and localizing language-aware keyframes seems a good idea to achieve this. - The method achieves strong results on a wide variety of benchmarks in both finetuning and zero-shot mode.

Weaknesses

- Most of the improvement compared to state-of-the-art methods come from the use of the BLIP-2 model. The proposed method only show small improvement over BLIP-2+concat in finetuning setting (see Table 1). - Can the method generalize with other image-language models, e.g. BLIP, and if so how does it perform? - The localizer relies on manual temporal language grounding annotations from QVHighlights. Could it benefit from weak supervision from e.g. HowTo100M like Moment-DETR?

Questions

See weaknesses.

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

Limitations and potential negative societal impact are discussed in the Supplementary Material.

Reviewer wzBa5/10 · confidence 4/52023-07-06

Summary

The paper presents the SeViLA framework, which addresses the limitations of existing image-language models for video question answering. It introduces two modules, Localizer and Answerer, that are fine-tuned from a pre-trained image-language model. SeViLA utilizes chaining for cascaded inference and self-refinement. The Localizer identifies language-aware keyframes, and the Answerer predicts the answer based on these keyframes. Additionally, the Answerer generates keyframe pseudo-labels to refine the Localizer, eliminating the need for costly video moment annotations.

Strengths

1. The paper is well-written and easy to understand. 2. The concept of localizing and then answering is intuitive and effective. 3. Extensive experiments demonstrate that SeViLA achieves state-of-the-art performance on challenging video question answering benchmarks. The paper also provides a comprehensive analysis, including comparisons with other localization models and the impact of keyframe variations.

Weaknesses

1. The motivation and LGDN (which also addresses VideoQA tasks) are quite similar. While the paper mentions LGDN in the introduction and related work sections (which I appreciate), I suggest the authors further clarify the differences between the two approaches in the related work section. 2. The self-chaining approach requires additional computational resources (key frames need to be recomputed). The paper should provide information on the additional computational costs involved.

Questions

Please see weaknesses.

Rating

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

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

2 fair

Limitations

Please see weaknesses.

Reviewer b5LJ6/10 · confidence 4/52023-07-10

Summary

The paper proposes a simple yet effective framework for video localization and question answering using pretrained image-language models. The key idea is to introduce a LOCALIZER module which localizes the keyframes in a video in order to ignore irrelevant frames and better answer the question. The LOCALIZER module is realized with a separate image-language model and the module can be further improved by fine-tuning on the pseudo labels generated by the QA results from the ANSWER module. The proposed framework, SEVILA, achieves state-of-the-art VQA results on multiple challenging benchmarks under both fine-tuned and zero-shot setups.

Strengths

1. The proposed framework is simple yet effective in extending image-language models to tackle video-related tasks such as video question answering and moment localization. The idea of using the LOCALIZER module to detect keyframes in a long video is technically sound and the results are convincing. 2. The paper provides extensive ablation study to analyze different components of the framework under different setups. E.g., the impact of selecting keyframes for VQA, the impact of self-refinement for the LOCALIZER module, and how the LOCALIZER compared with some existing keyframe localization methods. While there exists many keyframe localization methods, it's interesting to see the language-model-based method used in this paper gets comparable or even better results in the ablation study. 3. The paper is well organized, clearly written and the description of technical details is clear and easy to follow.

Weaknesses

1. The technical contribution of the work is relatively weak. The idea of selecting keyframes in a video for VQA is straightforward and widely explored in prior work. The proposed self-refinement strategy is also a standard semi-supervised learning approach and there's no new training strategies / techniques proposed for the fine-tuning of image-language models. 2. While I understand the main goal of this work is to utilize image-language models for VQA, the limitation of image-based models for video-related tasks should not be ignored. First, image-based models may not be sufficient to accurately localize the keyframes in a long video. Second, temporal modeling is also missing in the ANSWERER module, which makes the model less capable of modeling temporal dependencies compared with video-based models. These limitations may not be significant in the current VQA benchmarks since the questions are less sensitive to temporal dependencies and reasoning.

Questions

1. Is it necessary to use two different Q-former for the LOCALIZER and ANSWERER? Since both models are image based, what if we share the same Q-former for both modules? 2. What if we make the self-refinement an iterative process? In other words, we refine the ANSWERER module after the self-refinement of LOCALIZER, and vice versa. It's interesting to see when the performance would saturate with such interactive refinement process. A minor suggestion: I don't think Figure 4 is a good qualitative example because (1) the action of the ladies is barely visible in the small figure, let alone the intention of the action; (2) the question can be answered by only checking the question itself without knowing the visual content, since "staring out" and "hands above their eyes" are explicated mentioned.

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

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The authors have discussed the limitations of the work and its potential social impact.

Reviewer TABX5/10 · confidence 4/52023-07-11

Summary

This paper proposes a novel framework called Self-Chained Video Localization-Answering (SEVILA) that leverages a single image-language model (BLIP-2) for both temporal keyframe localization and question answering in videos. The framework consists of two modules, Localizer and Answerer, which are parameter-fine-tuned (LoRA finetuning) from BLIP-2. The paper demonstrates that the SEVILA framework outperforms several strong previous works in video question answering and event prediction benchmarks.

Strengths

(1) The paper proposes a novel framework that addresses the issue of missing important visual cues in video question answering tasks by leveraging a single image-language model for keyframe localization and question answering. (2) The self-chaining mechanism in the framework allows for cascaded inference and self-refinement, improving the accuracy of temporal localization without the need for expensive annotations. (3) The paper provides a comprehensive analysis of the proposed framework, including the impact of temporal localization, the self-refinement process, and the number of keyframes.

Weaknesses

1. In the backward chain, the pseudo-tags are frame-level, and if a single frame of information can not give the correct answer, the Localizer is considered to have given the wrong keyframe. This is obviously unreasonable and ignores the important time dependencies provided by the intermediate frame. This is probably why there was no significant change in their performance when the self-refine part was removed in the ablation study. 2. the proposal is not good at dealing with temporal causality. Because Blip-2 is a image-text MLLM model, there is no space-time inference, and their answer is to directly add the q-former query generated by each frame to the word token of LLM. This is equivalent to letting q-former himself to realize the time level of information understanding, and q-former does not have this ability. Could you make me clear?

Questions

see above.

Rating

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

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

3 good

Contribution

3 good

Limitations

see above.

Reviewer mWtS2023-08-10

I thank the authors for providing a rebuttal, have read the other reviews, and stand by my original rating of weak accept. Below are detailed answers: Weakness 1: I agree with the authors that adapting BLIP-2 to videos by concatenating visual features output from the Q-former outperforms applying BLIP-2 separately to all video frames and using voting. However, I do not see this as a major contribution given that concatenating visual features from different frames before multi-modal modeling is a standard technique used in video-text modeling, see MERLOT for instance. Weakness 2: I appreciate seeing that MiniGPT-4 also benefits from the localizer for zero-shot NeXT-QA, and encourage the authors to include additional image-text models and/or other evaluation datasets to confirm the generalizability of the proposed localizer+answerer approach. Weakness 3: I appreciate seeing that the NeXT-QA accuracy improves with the weakly-supervised pretraining of the localizer on ASR.

Authorsrebuttal2023-08-21

Thank you for your thoughtful feedback and we are glad that you appreciate MiniGPT-4 and ASR results on NeXT-QA. We agree that concatenating visual features isn’t our primary contribution (although one interesting difference is that we do not require fine-tuning). We will add all these changes and clarifications to the final version.

Area Chair kEmK2023-08-17

Please respond to authors' rebuttal

Dear Reviewers, Thanks for your contributions! Please respond to the authors after carefully reading their rebuttals and other reviews. If your assessment of the paper changes, please update your score with a short justification for the new rating. Thank you, AC

Reviewer b5LJ2023-08-21

Thanks for your response. Most of my concerns are well addressed in the rebuttal and I'd keep my original positive rating.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC