acceptodds
Desk rejected at ICLR 2027

Seeing Anomalies in the Dark: Video Anomaly Understanding with Low-Light Conditions for Multimodal LLMs

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Video anomaly understanding (VAU) aims to detect and classify anomalous events in videos, providing fine-grained explanations of their nature and context. However, existing works in this direction often assume well-lit scenes, paying limited attention to low-light scenarios where many real-world anomalies occur, e.g., nighttime intrusions and blackout looting. To address this challenge, we introduce VAULT, a benchmark for low-light video anomaly understanding across various granularities. Specifically, VAULT contributes: (1) broad coverage of poorly illuminated anomalies with four illumination levels, covering 25 anomaly categories across 6 scene types, and (2) a unified evaluation suite comprising four tasks, i.e., temporal grounding, anomaly classification, anomaly description, and video question answering (VQA). Building upon VAULT, we propose EventVAULT, an event-guided framework that combines RGB appearance cues with event-stream dynamics for low-light VAU. Instead of relying solely on RGB appearance, EventVAULT leverages event streams to capture transient dynamics and address three key challenges: when to acquire informative RGB evidence, where to focus within the scene, and how anomalous events evolve. Extensive experiments demonstrate that the performance of existing MLLMs deteriorates as illumination decreases, while EventVAULT consistently improves performance across all four tasks. Our dataset and findings jointly highlight the need for more robust VAU under extreme visual conditions. Code available at: https://anonymous.4open.science/r/VAULT.

Primary area:
datasets and benchmarks
Keywords:
Video Anomaly Understanding, Multimodal Large Language Model

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_QeoWOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces VAULT, a benchmark for low-light video anomaly understanding containing 1,776 real-world videos across four illumination levels. It supports temporal grounding, anomaly classification, anomaly description, and video question answering. The authors also propose EventVAULT, which uses synthetic event streams to guide RGB frame sampling, localize informative regions, and provide motion cues to multimodal LLMs. Experiments evaluate existing models on VAULT and demonstrate improvements with the proposed framework.

Critical strengths
  • The motivation and the organisation of the paper is clear.
  • The feature-space analysis provides useful evidence that normal and anomalous frames are less separable under lower illumination, offering insight into the challenges of low-light anomaly understanding and motivating the proposed approach.
  • VAULT is a benchmark that addressing an underexplored setting. VAULT focuses on low-light video anomaly understanding for multimodal LLMs, with 1,776 real-world videos spanning 41.37 hours, 25 anomaly categories, and six scene types across indoor and outdoor environments.
  • The benchmark supports four complementary tasks - temporal grounding, anomaly classification, anomaly description, and video question answering and groups videos into four illumination levels, enabling evaluation of multiple capabilities across lighting conditions.
Critical weaknesses
  • The introduction does not clearly define ''illumination level'' and ''RGB motion observability''. Please specify the metrics used to quantify each and the thresholds used to divide samples into groups in Figure 1.

  • The claim of being the first benchmark dedicated to low-light video anomaly understanding requires justification. In particular, the authors should discuss A2Seek [a] and ReCI [b], which supports anomaly understanding and explicitly considers low-light and nighttime scenarios, and clarify how VAULT differs in its focus and evaluation protocol. Otherwise, the novelty claim may be overstated.

  • The novelty of event-guided temporal and spatial evidence selection is insufficiently established. EventSTU [c] already uses events for RGB keyframe sampling and spatial token pruning in Video-LLMs, while event-guided RGB cropping has also been explored in EPCNet. The authors should clarify the technical differences and demonstrate the benefits of TEA and SEL over closely related strategies.

  • Given that EventSTU is closely related, the authors should consider comparing against it under matched experimental settings.

  • Video examples are not provided, making it difficult to assess the dataset’s visual quality and the challenges posed by its low-light conditions.

[a] Mo, M., Tong, X., Tan, M., Leng, J., Zheng, J., Liu, Y., Chen, H., Gan, J., Li, W. and Gao, X., 2026. A2seek: Towards reasoning-centric benchmark for aerial anomaly understanding. Advances in Neural Information Processing Systems, 38.

[b] Mo, M., Zheng, J., Leng, J. and Gao, X., 2026, June. Retrieval-Guided Contextual Inference for Training-Free Video Anomaly Detection in Low-Light Scenarios. In Proceedings of the 2026 International Conference on Multimedia Retrieval (pp. 930-939).

[c] Xu, W., Dong, X., Li, Y., Shi, H. and Xiong, Z., 2025. EventSTU: Event-Guided Efficient Spatio-Temporal Understanding for Video Large Language Models. arXiv preprint arXiv:2511.18920.

Questions for authors
  • Can the authors provide visual examples of TEA/SEL’s frame selection and spatial crops when irrelevant motion is stronger than the anomalous motion? Such examples would help assess whether the method preserves anomaly-relevant evidence despite distracting event activity.

  • How does TEA/SEL handle crowded scenes with complex, simultaneous human motions, such as retail stores where subtle theft may be overshadowed by normal customer activity? It would be great that the author can provide visualizations showing whether the selected frames and spatial crops retain the anomaly-relevant evidence in such cases.

1 comment

I thank the authors for their detailed responses and additional experiments. The rebuttal has addressed my concerns regarding novelty, clarity, and experimental comparisons. I am therefore happy to raise my rating to Weak Accept.

Reviewer_9S59OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces VAULT, a benchmark for video anomaly understanding under natural low-light conditions, containing 1,776 videos and covering localization, classification, description, and question answering. EventVAULT synthesizes events from RGB video for frame selection, region cropping, and dynamic representation, and then uses RGB frames and event images for instruction tuning of a multimodal model. Validation includes comparisons across multiple models, component ablations, and tests on three external datasets.

Critical strengths
  1. VAULT combines videos captured under natural low-light conditions with localization, classification, description, and question answering, providing a dedicated evaluation resource for studying anomaly understanding when visual evidence is degraded. This is the paper’s most valuable contribution.

  2. In the LoRA comparison using the same Qwen3-VL-8B backbone, localization mIoU and VQA accuracy improve by 7.00 and 4.70 percentage points, respectively. The component ablations further examine the roles of TEA, SEL, and AEER, supporting the effectiveness of this approach to organizing visual inputs.

Critical weaknesses
  1. The events are synthesized from RGB, and the paper needs to explain their added value over representations such as frame differences and optical flow. The localization results with TEA alone are already close to those of the full model, but the experiments do not compare these dynamic inputs using the same backbone and visual input budget. The current results have not established the comparative advantage of event images, which directly affects the assessment of the method’s incremental contribution.

  2. All three modules rely on event activity to allocate observations, but strong brightness changes do not necessarily correspond to important anomaly cues. The paper does not report EventVAULT’s results relative to RGB grouped by the strength of anomalous motion and background motion, and the data filtering also excludes videos dominated by severe camera shake, flashing, or similar interference. The existing evaluation does not show whether this selection strategy can retain useful evidence when prominent motion comes from the background and the key anomalous action involves little motion.

  3. Video content is not held constant across illumination groups, so the effects of illumination cannot be separated from those of scenes and event content. The external tests also do not report the RGB results corresponding to EventVAULT under the same VAULT training conditions. Consequently, the paper cannot yet clearly attribute these gains to design choices specifically targeting low light or to transfer advantages provided by the event module.

Questions for authors
  1. Do the AEER ablations keep the sampled RGB frames, cropping results, and total visual input budget fixed? Are there results comparing event images with frame differences and optical flow under the same conditions that show the gains come from the event representation itself?

  2. For samples where anomalous actions involve little motion or background motion is strong, does EventVAULT still improve over the RGB baseline? Please provide results for the relevant groups and explain whether frame selection and cropping retain the key evidence that determines the anomaly category.

  3. Have the illumination groups been matched by category and scene, and what are EventVAULT’s gains over RGB after matching? Which backbone and trained weights are used in the external evaluation?

Reviewer_mwuXOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates video anomaly understanding (VAU) under low-light conditions. The authors introduce VAULT, a benchmark containing 1,776 real-world videos across 25 anomaly categories and four illumination levels, supporting temporal grounding, anomaly classification, description, and VQA. The authors further propose EventVAULT, an event-guided framework that integrates temporal evidence allocation (TEA), spatial evidence localization (SEL), and adaptive event evidence representation (AEER) to enhance RGB-based anomaly understanding. Experiments across multiple MLLMs, along with ablation and cross-dataset evaluations, demonstrate improvements in VAU performance, particularly in temporal grounding and VQA.

Critical strengths
  1. A valuable benchmark for low-light video anomaly understanding. The proposed VAULT benchmark addresses an underexplored yet practically important problem by providing real-world low-light videos with comprehensive annotations for four VAU tasks. Its diverse illumination conditions enable systematic evaluation of MLLM robustness, making it a useful resource for future research.

  2. An effective event-guided framework with comprehensive empirical validation. EventVAULT integrates temporal sampling, spatial localization, and event-based dynamic representations to improve anomaly understanding. The reported improvements across multiple MLLMs, supported by ablation studies and cross-dataset evaluations, demonstrate the practical potential and general applicability of the framework.

Critical weaknesses
  1. Insufficient clarification of motion observability and the applicability of event-guided evidence selection. In Figure 1 (Left), the authors analyze model performance across different levels of RGB motion observability, but the definition and measurement procedure of this metric are unclear. It is particularly important to clarify whether motion observability is objectively computed or subjectively annotated, as subjective assessments could affect sample grouping and the resulting performance analysis. Additionally, TEA and SEL in Section 4.2 primarily rely on event activity, which may not always correspond to anomaly-relevant evidence in scenes with multiple moving objects or substantial background changes. A clearer description of the metric, along with representative failure cases or a discussion of the method's applicability, would strengthen the analysis.

  2. Limited analysis of feature separability and the role of event-based representations. In Figure 1 (Middle), the authors analyze the separability between normal and anomalous frame features under different illumination conditions using the pretrained Qwen3-VL-8B-Instruct model. However, no corresponding analysis is provided after fine-tuning. Comparing feature separability between the RGB-only SFT baseline and EventVAULT could provide further insights into whether the method improves anomaly-discriminative representations. Additionally, Table 3 shows that TEA alone contributes substantially to the temporal grounding improvement. A brief discussion or supplementary comparison with conventional motion cues, such as frame differences or optical flow, would help clarify the specific benefits of event-based guidance.

Questions for authors
  1. Clarification of RGB Motion Observability (Figure 1, Left). How is RGB Motion Observability defined and measured in Figure 1 (Left)? Is it computed using an objective metric or assessed through human annotations? If subjective assessments are involved, how is their consistency ensured? Clarifying the measurement and grouping criteria is important for assessing the reliability and reproducibility of the reported analysis.

  2. Feature Separability and the Role of Event Representations (Figure 1, Middle; Table 3). Figure 1 (Middle) analyzes feature separability using the pretrained Qwen3-VL-8B-Instruct model. Does fine-tuning with EventVAULT further improve the separability between normal and anomalous representations, particularly under low-light conditions? Additionally, given the substantial improvement achieved by TEA alone in Table 3, could the authors clarify the respective contributions of evidence selection and event-based representations to the overall performance gains?

  3. Applicability of Event-Guided Evidence Selection (Section 4.2). Since TEA and SEL primarily rely on event activity, how does the framework handle scenes where irrelevant background motion or illumination changes dominate the event responses? Could the authors discuss representative failure cases or clarify the conditions under which event-guided selection may become less reliable? This would help assess the practical applicability and limitations of the proposed method.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.