MECD: Unlocking Multi-Event Causal Discovery in Video Reasoning

Video causal reasoning aims to achieve a high-level understanding of video content from a causal perspective. However, current video reasoning tasks are limited in scope, primarily executed in a question-answering paradigm and focusing on short videos containing only a single event and simple causal relationships, lacking comprehensive and structured causality analysis for videos with multiple events. To fill this gap, we introduce a new task and dataset, Multi-Event Causal Discovery (MECD). It aims to uncover the causal relationships between events distributed chronologically across long videos. Given visual segments and textual descriptions of events, MECD requires identifying the causal associations between these events to derive a comprehensive, structured event-level video causal diagram explaining why and how the final result event occurred. To address MECD, we devise a novel framework inspired by the Granger Causality method, using an efficient mask-based event prediction model to perform an Event Granger Test, which estimates causality by comparing the predicted result event when premise events are masked versus unmasked. Furthermore, we integrate causal inference techniques such as front-door adjustment and counterfactual inference to address challenges in MECD like causality confounding and illusory causality. Experiments validate the effectiveness of our framework in providing causal relationships in multi-event videos, outperforming GPT-4o and VideoLLaVA by 5.7% and 4.1%, respectively.

Paper

Similar papers

Peer review

Reviewer 17Ce8/10 · confidence 2/52024-07-04

Summary

The paper studies causal reasoning in video, specifically, causal diagrams in long, multi-event videos. To do this, the authors attempt to define MEDC, a new task for discovering the complete causal relation diagram in multi-event chronological videos (~ 2 minutes). For example, analyzing events from traffic surveillance videos across different times, to identify the causes of an accident. Coupled with the task, they create a dataset of 1107 lifestyle videos from ActivityNet with multiple events in them. To facilitate the task, they propose Video Granger Causality Model, a framework inspired by the Granger Causality Method, and show it surpasses existing models (GPT-4, Gemini-1.5 Pro, Video-LLava).

Strengths

* The paper introduces a new, challenging, and important task in video understanding * The methodology section was interesting to read, with the approach being mostly easy to follow. * There is a wealth of analyses and ablation work both in the main paper and the supplementary, which cover many aspects of the paper, such as the task itself, the trained model, and the baselines the model is compared to.

Weaknesses

No major weaknesses come to mind.

Questions

N/A

Rating

8

Confidence

2

Soundness

4

Presentation

4

Contribution

3

Limitations

yes

Reviewer 2Vef5/10 · confidence 4/52024-07-12

Summary

This paper introduces a new task called Multi-Event Causal Discovery (MECD). Given a video which comprises multiple temporal events that are chronologically organized, the goal is to predict if any previous event has a causal effect on the last event in the sequence. The MECD dataset is filtered and curated from the existing ActivityNet v1.3 video dataset. Additionally, the paper also proposes the Video Granger Causality Model (VGCM) approach, which leverages an event prediction model inspired by the Granger Causality method to perform an Event Granger Test. This test estimates causality by comparing predicted result events with masked versus unmasked premise events. Furthermore, the front-door adjustment and counterfactual inference techniques are also incorporated into VCGM. The authors demonstrate the benefits of their proposed VCGM approach by comparing it to state-of-the-art LLMs and VLLMs on the MECD benchmark, where it outperforms the latter by a significant margin.

Strengths

1) Overall, the paper addresses a significant problem in video understanding, especially with the recent interest in understanding the capabilities of visual LLMs (VLLMs). In contrast to how existing visual-language foundation models are pretrained and evaluated, the proposed MECD benchmark and the Video Granger Causality Model are aimed at understanding the causal relationships between multiple temporal events in videos instead of simply describing them. 2) The intuition and theoretical reasoning underlying the proposed Video Granger Causality Model is sound. In particular, the two main problems of Causality confounding and Illusory Causality in video causal discovery are well-motivated. Additionally, the mathematical formulation for integrating the front-door adjustment and counterfactual inference methods to resolve the listed issues is insightful. 3) The model figures are informative and especially helpful in helping the reader to understand the different stages of the data curation process as well as the intuition behind each stage. The paper is also well-organized and well-written.

Weaknesses

1) While the introduced Multi Events Causal Discovery dataset is a nice contribution as a benchmark, it is relatively limited in size with 1107 data samples. For example, there are only 808 and 299 video samples for training and evaluation, respectively. With such a small number of evaluation samples, it is challenging to evaluate these trained LMMs comprehensively. It may be beneficial to source videos from a wider variety of sources such as long and instructional videos from Youtube. Additionally, why is a validation set not included as part of the splits? 2) The results in Table 1 provide some interesting comparisons between language-only baselines like Gemini 1.5 Pro and VLLM baselines including Minigpt4-video and Video-LLaVA. However, there are some other state-of-the-art video-language models that are not compared to in the paper such as Video-Llama [1]. Additionally, it has been demonstrated before that the image variants of VLLMs actually outperform their video counterparts on some video understanding tasks. It may be beneficial to include comparisons to such models such as InstructBLIP [2] and Llava to make the analysis more comprehensive. 3) In section 4.1, it would be helpful to include a brief overview of the Granger Causality method, which forms the basis of the proposed Video Granger Causality Model. Additionally, there are some sections, especially those that briefly describe the Event Granger Test and the causal inference techniques, could benefit from additional technical depth and explanations. 4) Another concern lies in the training and evaluation setup. The VGCM approach is built off the VideoBert model and it will be helpful to include the results obtained by the base VideoBert model. Furthermore, the other baselines, that are being compared to, are evaluated under a zero-shot setting. However, the proposed model appears to be trained and evaluated under a strongly-supervised setting with the annotations from the MECD dataset. [1] Hang Zhang et al. Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding. EMNLP 2023 demo track. [2] Wenliang Dai et al. InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning. NeurIPS 2023.

Questions

This is a minor note but it will be helpful for readers to define the term ‘premise event’ at the beginning. Also, please look at the above-mentioned limitations.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes.

Reviewer eXFX5/10 · confidence 4/52024-07-12

Summary

The paper introduces a new task and dataset, Multi-Event Causal Discovery (MECD) to better understand long videos in a casual perspective. Inspired by the Granger Causality method, the authors devise a framework, dubbed VGCM, to perform the Event Granger Test. Also, VGCM is combined with front-door adjustment and counterfactual inference to tackle the issues of causality confounding and illusory causality. Experiments show that VGCM outperforms GPT-4 and Video-LLaVA in providing causal relationships in multi-event videos.

Strengths

- To advance comprehensive and structured causality analysis for videos with multiple events, the authors introduce a benchmark for the multi-event causal discovery task. - The authors have crafted an innovative model framework that integrates the Event Granger Test with various causal inference techniques. - Experiments demonstrate the efficacy of the proposed framework in providing causal relationships within multi-event videos.

Weaknesses

- To enhance the demonstration of the model's generalizability, it is recommended to conduct experiments on a variety of related datasets. - In Figures 1(c) and 1(e), the events occurring within the video frames are not readily discernible. Additional verbal descriptions are recommended to facilitate a clearer understanding of the figures' intended message.

Questions

- The authors employ Accuracy as the primary performance metric. Could the authors consider or design some other metrics that are more suitable for Multi-Event Causal Discovery (MECD)? - It would be highly beneficial if the dataset and associated code could be made publicly available at the earliest opportunity. - Regarding lines 154-155, I suspect that the confounding factor in the spurious causal relationship could introduce significant discrepancies when comparing the output features with and without the factor. Could the authors provide further analysis on this matter?

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The model depends largely on video captions, which makes it possibly incapable of processing video datasets without captions or descriptions.

Reviewer 9my45/10 · confidence 4/52024-07-13

Summary

The paper introduces the Multi-Event Causal Discovery (MECD) task, aiming to uncover causal relationships in videos with multiple events. It presents a novel framework inspired by the Granger Causality method, utilizing a mask-based event prediction model to perform causal inference. The paper also introduces a new dataset for training and evaluation, demonstrating the framework's effectiveness in outperforming existing models like GPT-4 and Video-LLaVA.

Strengths

- Novelty and Relevance: The paper addresses a significant gap in video reasoning tasks by focusing on multi-event causal discovery, which is more reflective of real-world scenarios. - Framework Design: The use of Granger Causality, combined with advanced causal inference techniques like front-door adjustment and counterfactual inference, is innovative and well-justified. - Empirical Validation: The framework demonstrates substantial performance improvements over state-of-the-art models, indicating the robustness and efficacy of the proposed approach. - Dataset Contribution: The creation of the MECD dataset, with detailed annotations and diverse scenarios, is a valuable contribution to the field.

Weaknesses

- Complexity of Implementation: The framework's reliance on multiple advanced techniques may pose challenges for implementation and replication by other researchers. Besides, does VGCM run slower than those VLLM baselines? - Missing Details: What's the architecture of different encoders? How do the authors pretrain them? How do different loss design? The authors should give more details about fair comparisons. - Some Typos: - In Figure 3, the box colors for the masked and unmasked input should be green and orange. - In Line 166, it should be $Env_{V}$ and $Env_{C}$.

Questions

I appreciate that the authors proposed a more novel and interesting setting. However, given the missing details and overly complicated pipeline, I tend to reject this paper for now. And I hope the authors can provide more details about the training and architecture. This will help us reviewers to make a fair evaluation.

Rating

5

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

Yes.

Authorsrebuttal2024-08-13

Response to reviewer-9my4

Thanks for your great efforts and time in reviewing our paper. ### 1. *Results of fine-tuned Video-LLaVA* Firstly, we want to state that the few-shot evaluation (In-Context Learning) for LLMs is widely recognized as a strong baseline for reasoning and causal discovery tasks [1-6], effectively reflecting their performance on downstream tasks. Therefore, comparing LLMs in a few-shot setting is a common and accepted practice in the field. Secondly, neither GPT-4 nor Gemini-Pro offer interfaces for fine-tuning, consequently, we reported the few-shot results of LLMs and investigated the impact of varying the number of few-shot examples, demonstrating the adequacy of our prompt (cf. supplementary Sec. G, lines 572-578). Furthermore, as you and reviewer-2Vef suggested, to ensure a fairer comparison, we conducted a strongly-supervised experiment on the open-source method Video-LLaVA. Specifically, we fine-tuned Video-LLaVA using LoRA under its official implementation on our MECD training set. As shown in the table, Video-LLaVA gains a 4.6% improvement from fine-tuning on our MECD. **However, it still falls behind our proposed method**. Since fine-tuning LLM-based baselines is time-consuming, we will include more results of fine-tuning VLLMs for a comprehensive comparison in our final version. |Setting| Acc| |------------------------ | -------- | |Video-LLaVA (few-shot)| 62.5| |Video-LLaVA (fine-tuned) |67.1| |Ours|**71.2**| [1] Language models are few-shot learners. NeurIPS 2020. [2] Rethinking the role of demonstrations: what makes in-context learning work? EMNLP 2022. [3] Large language models are latent variable models: Explaining and finding good demonstrations for in-context learning. NeurIPS 2023. [4] CLADDER: assessing causal reasoning in language models. NeurIPS 2023. [5] Causal inference using llm-guided discovery. AAAI 2024. [6] Open Event Causality Extraction by the Assistance of LLM in Task Annotation, Dataset, and Method. ACL 2024. ### 2. *More details* We're happy to provide more details here to clarify your concerns. However, we need to state that **these details can be also found in the supplementary in our initial submission**. To further relieve your concerns about our implementation and reproducibility, we provide an **anonymous GitHub project** (https://anonymous.4open.science/r/NeurIPS-4887-MECD), which contains our *data annotation*, *training and evaluation codes*, and details about how we evaluate LLMs and how we fine-tuning the Video-LLaVA. Additional details can be found in the README.md file. - **Training procedure**: All the experiments are conducted on 1 NVIDIA A40 GPU. We train our model for 20 epochs with a learning rate of 16e-5 about 6 hours. Our optimizer is consistent with the BertAdam optimizer, with 3 epochs of warm-up. We report the average results during all experiments under three random seeds (2023, 2024, 2025). - **Hyperparameters**: $\lambda_C$, $\lambda_R$, $\lambda_V$,$\lambda_S$ are set to be $1.0, 4.0, 0.25, 0.05$. - **Encoder & Decoder architecture**: Our encoder $Enc_{v}$, $Enc_{c}$, and multi-modal video decoder $Dec$ are built upon videoBERT, a joint model for video and language representation learning. - **Evaluation of LLMs**: We prompt the GPT-4 with the following few-shot prompts to conduct evaluation (Details can be found in the anonymous GitHub project): ``` # Task: Each video consists of n events, and the text description of each event has been given correspondingly (separated by " ",). You need to judge whether the first n-1 events in the video are the cause of the last event, and the probability of the cause 0(uncausal) or 1(causal) is expressed as the output, Let's think step by step through the chain of thought. Here are several examples of judging whether the first n-1 events in the video are the cause of the last event: <start> First example: Text description of n events: ... The probability output should be: ... Second example: Text description of n events: ... The probability output should be: ... Third example: Text description of n events: ... The probability output should be: ... <end> ``` As a long-term maintenance benchmark, we confirm that the dataset and code will be public after being accepted. We hope our responses can address your concerns and raise your rating of our paper.

Reviewer 9my42024-08-14

Response to authors

Thanks for the further claim, and I slightly raise the rating. I suggest adding these additional results and details in the final version to make the whole paper more reliable and reproducible.

Authorsrebuttal2024-08-14

Response to reviewer-9my4

Thanks for your response. We really appreciate your acknowledgment that our further claim effectively addressed your concerns. We will provide the additional results and all the details in the final version of the manuscript following your suggestion. Besides, we will make the complete dataset and codes available as soon as the manuscript is accepted. Once again, we would like to express our sincere gratitude for your positive feedback and for contributing to our manuscript.

Authorsrebuttal2024-08-12

Kind Request for Discussions and Feedback for Paper 4887

Dear Reviewers, Firstly, we'd like to express our sincere gratitude for taking the time to review our paper and for providing invaluable feedback. Since the deadline for discussion is just around the corner, if you still have any concerns about our paper after reading our response, please feel free to contact us by adding comments. Your engagement during this discussion period is crucial for the improvement of our work. We genuinely value your insights and look forward to your continued feedback. Thank you for your time and consideration. Best regards, Authors of Paper 4887

Reviewer eXFX2024-08-12

Thank you for your detailed response. My concerns have been addressed, and after considering the feedback from other reviewers, I will maintain my current rating.

Authorsrebuttal2024-08-12

Thanks for your response. We greatly appreciate your acknowledgment that our rebuttal has effectively addressed your concerns. We have duly noted your suggestion regarding the model's generalizability and the additional metric in the final version of the manuscript. We will ensure that this change is made to improve our work. Once again, we would like to express our gratitude for your valuable feedback and for contributing to the improvement of our manuscript.

Reviewer 9my42024-08-13

Response to authors

The authors solve part of my issues. However, some key issues are not clear, like the details to reproduce the whole results. It will be better if the authors can provide detailed training hyperparameters, architecture, data in clear tables, and even the code. Besides, the LLM-based models are not fine-tuned with the proposed MECD, which is unfair. Can the authors provide more fair results with training on VideoChat or Video-LLaVA. And the authors should clearly claim they use `few-shot paradigm` in the main text. Considering the problems, I maintain my original rating.

Reviewer 2Vef2024-08-13

Thank you very much for your detailed responses to my questions. In particular, I find the addition of the new Ave SHD metric to be helpful and especially insightful for understanding the limitations of existing large multimodal models. Consequently, I will retain my initial rating.

Authorsrebuttal2024-08-13

Response to reviewer-2Vef

Thank you for your response. We appreciate your acknowledgment of the usefulness of our rebuttal. We confirm that we will include the additional SHD metric and improve the comparison methods as you suggested in the final version of the manuscript. Once again, we are grateful for your valuable feedback and acknowledgment of our work.

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC