Generalizing Single-Frame Supervision to Event-Level Understanding for Video Anomaly Detection

We deeply appreciate the reviewer's meticulous assessment and profound insights. The positive recognition of our novel supervision paradigm, effective two-stage framework, and dataset contribution is especially encouraging. We're also thankful for the constructive criticism, which is invaluable for refining our manuscript. Here are our responses to the concerns:

Response to Q1: Multi-event Analysis

For multi-event generalization validation, class-wise performance on XD-Violence is further evaluated, where more complex behaviors, e.g., Riot, and interfering scenarios are involved.

Notably, our performance steadily surpasses that of the SOTA weakly-supervised methods across diverse abnormal classes, which aligns with the evaluation on UCF-Crime:

These results further confirm the generalizability of FPL across varied abnormal behaviors. With the combination of dynamic anomaly relevance estimation and static feature similarity evaluation, diverse abnormal intervals are mined, which ultimately enables accurate capture of the essence of abnormal behaviors even in diverse and noisy scenarios.

Response to Q2: Computational Overhead

For practical usability evaluation, model parameters and inference time are further investigated:

Remarkably, our SF-VAD method demonstrates a smaller parameter size, much faster inference time, and better detection performance than previous state-of-the-art semi-supervised and weakly-supervised VAD methods. The superior performance-cost trade-off stems from the precise abnormal learning target enabled by FPL, which allows a lightweight network architecture to effectively capture diverse anomalous cues, without computationally expensive memory modules and text encoders.

Response to Q3: More Complete Baseline Comparison

Due to the page limitation in initial submission, the semi-supervised methods are not included. As shown in Sec. H in Supp., we demonstrate more comprehensive comparison results with the state-of-the-art semi-supervised methods including, MULDE[1], AED-MAE[2], MGEnet[3], LANP[4], etc. While we cannot locate the specific SWAD literature, if the reviewer could kindly provide it, we would be delighted to perform this comparison and include the appropriate citation in the revised manuscript. This would further illustrate our method's broad applicability and effectiveness.

The complete evaluation results will be included in the main manuscript in the revision.

Response to Q4: Intuitive Justification for Evidential Learning

As mentioned in line 150-152, feature similarity itself is not fully dependable due to its insensitivity to subtle changes in local spectrum. Therefore, evidential learning is introduced, which provides uncertainty in its predictions as an estimation of the discrepancy from annotated anomalies. Specifically, FPL first employs an evidential learning model to learn noise-free abnormal patterns using the reference of SF-VAD solely. During the subsequent stage, evidential learning model predicts uncertainty to measure the deviation between input distribution and modeled anomaly, which serves as a proxy for anomaly relevance. Via the relevance, we generalize the anomaly supervision from a single frame into more complete scope. For a detailed explanation:

Why Evidential Learning: In comparison to standard probabilistic models, evidential learning estimates the deviation from the learnt accurate abnormal patterns by prediction uncertainty. In contrast, standard probabilistic models only predict certain anomaly probability, even when facing unseen inference data. Therefore, evidential learning facilitates a reliable generalization from a single frame to the complete abnormal interval through uncertainty estimation, which probabilistic models fail to do.

How Evidential Learning Works: Evidential learning models the distribution of probabilities by assigning evidence to support predictions, rather than directly estimating probabilities. In our SF-VAD framework, we specifically apply this by fitting a Beta distribution to model anomaly evidence. The parameters of this Beta distribution directly quantify the strength of our belief in each prediction. Critically, this approach allows us to not only derive a point estimate of anomaly probability but also explicitly quantify the uncertainty associated with that prediction.

Evidential Learning's Advantages: For epistemic uncertainty [5] estimation, previous methods [6, 7] adopt Bayesian Neural Network or Monte Carlo Dropout. Yet, these methods require multiple inferences to deduce the uncertainty and incur excessive computational cost. In contrast, evidential learning estimates uncertainty based on the Theory of Evidence and predicts both probability and uncertainty with a single forward pass, ensuring computational efficiency.

Response to Q5: Event-wise Robustness Test

For event-wise robustness test, we provide a performance comparison for video clips containing different numbers of disjoint abnormal events in XD-Violence.

Regarding the request for UCF-Crime, we unfortunately find that the test set of UCF-Crime contains at most two disjoint abnormal events per abnormal video. Therefore, we are unable to assess AUC scores for video clips with three or more disjoint abnormal events as specifically requested by the reviewer due to this inherent dataset limitation. Still, to further demonstrate our method's robustness to varying numbers of abnormal events, we conduct additional experiments on UCF-Crime with the available event counts.

FPL consistently outperforms baseline models under weak supervision across diverse numbers of abnormal events. It further verified the effectiveness of FPL in generalizing single-frame supervision into diverse abnormal events.

Response To Q6: Robustness to Non-ideal Annotations

To evaluate FPL's effectiveness under low-quality or inconsistent single-frame annotation, we further evaluate the performance of FPL using annotations from GlanceVAD [8] and ARG [9]. Fixing the hyperparameters, we randomly choose frames from glance annotation or a frame within the abnormal interval by the full annotations. The results are shown below.

As shown in the table above, the performance of FPL remains consistent under different single frame settings, which demonstrates a good generalizability under varying quality of annotations. The superior generalization performance comes from the dynamic abnormal event reasoning, where noise-free abnormal patterns are first modeled, then the precise abnormal intervals are mined based on the relevance and feature similarity to the learnt abnormal behaviors. In contrast to previous methods [10], which model the relative position within the abnormal interval by a fixed distribution, our method is much more position-agnostic, as abnormal intervals are dynamically deduced. This ensures excellent generalization capabilities, even in scenarios with ambiguous or non-representative single-frame annotations.

[1] Micorek, Jakub, et al. "Mulde: Multiscale log-density estimation via denoising score matching for video anomaly detection." CVPR, 2024.

[2] Ristea, Nicolae-C., et al. "Self-distilled masked auto-encoders are efficient video anomaly detectors." CVPR, 2024.

[3] Yang, Guoqing, et al. "A multilevel guidance-exploration network and behavior-scene matching method for human behavior anomaly detection." ACM MM, 2024.

[4] Shi, Haoyue, et al. "Learning Anomalies with Normality Prior for Unsupervised Video Anomaly Detection." ECCV, 2024.

[5] Hüllermeier, Eyke, and Willem Waegeman. "Aleatoric and epistemic uncertainty in machine learning: An introduction to concepts and methods." Machine learning, 2021.

[6] Subedar, Mahesh, et al. "Uncertainty-aware audiovisual activity recognition using deep bayesian variational inference." ICCV, 2019.

[7] Zhang, Chen, et al. "Exploiting completeness and uncertainty of pseudo labels for weakly supervised video anomaly detection." CVPR, 2023.

[8] Zhang, Huaxin, et al. "Glancevad: Exploring glance supervision for label-efficient video anomaly detection." arXiv, 2024.

[9] Liu, Kun, and Huadong Ma. "Exploring background-bias for anomaly detection in surveillance videos." ACM MM, 2019.

[10] Cui, Ran, et al. "Video moment retrieval from text queries via single frame annotation." SIGIR, 2022.

We sincerely appreciate the reviewer's meticulous feedback and profound insights into our method. Meanwhile, we are also grateful for the recognition of SF-VAD's effectiveness, experimental completeness, and clarity of writing. To address the concerns, we'd like to offer the following clarifications：

Response to Q1: Justification of SF-VAD

Advantages of SF-VAD: SF-VAD significantly optimizes the utility of limited anomalous samples. On the one hand, it is highly annotation efficient and accurate . Since annotators must already perceive an obvious abnormal frame to label an entire video as abnormal. The additional annotation time is almost negligible:

Simultaneously, compared to fully-supervised methods, it avoids the inherent ambiguity of anomaly boundaries, resulting in remarkable annotation precision.

On the other hand, SF-VAD delivers precise anomaly reference, which is absent in the weakly-supervised paradigm, facilitating evidential and comprehensive anomaly learning and resulting in promising detection performance:

Learning paradigm of FPL: Leveraging single frame annotation, FPL effectively mines evidential abnormal intervals via integration of dynamic anomaly relevance reasoning and static feature similarity, abandoning the groundless and notorious top-K selection in MIL. As a result, a much clearer abnormal learning target is revealed, facilitating a more discriminative classification boundary.

Response to Q2: Model's Robustness to Shifting Annotations

Authentic Annotation Protocol: In the proposed three benchmarks, the annotators are allowed to freely navigate the video timeline and locate a random abnormal frame. Such annotation protocol ensures that the annotation reflects genuine human interests in realistic settings without any prior guidance that could artificially inflate performance. Details can be found in Sec. B in Supp.

Robustness to single-frame annotation bias: To further validate the robustness of our method under varying single-frame annotation, we conduct additional experiments by sampling two set of abnormal frames from GlanceVAD [1] and ARG [2] in UCF-Crime:

The performance of FPL remains mostly consistent with randomly selected ambiguous or non-representative single-frame annotations. In contrast to previous methods [1, 3] with a fixed distribution, our method is much more position-agnostic, as abnormal intervals are dynamically deduced.

Response Q3: Effectiveness in Multi-event Scenarios

We appreciate the reviewer for highlighting multi-event issues, which are pivotal for realizing robust anomaly detection. Specifically, we tailor an abnormal event mining algorithm for multi-interval anomaly mining, which integrates anomaly relevance and feature similarity. This algorithm first selects potential abnormal frames by assessing the similarity across both action and appearance features, with the selection further refined by variance verification and relative gap constraints to enhance exploration accuracy. It then utilizes these pending frames to gather similar frames into the intervals by estimating anomaly relevance. For validation, we present ablation experiments on multi-event anomalies in XD-Violence:

For further evaluation, we extend our evaluation to include multi-event experiments on UCF-Crime. The results are as follows:

These experimental results further substantiate our method's generalizability in capturing patterns from multiple abnormal segments, which is primarily attributed to multi-event anomaly mining algorithm.

Response to Q4: Investigation on Generalization

For a comprehensive generalization evaluation, we consider the following six aspects, ordered by importance:

Performance Consistency across Various Domains: it directly measures the model's generalizability across different real-world distributions. As shown in Tab. 1 and Tab. 2, our method consistently outperforms SOTA methods with higher AUC and lower False Alarm Rate across a wide spectrum of domains, anomaly behaviors, perspectives, and data sources.

Detection of Diverse Anomaly Types: we provide a class-wise AUC analysis on UCF-Crime:

In addition, we further extend our evaluation to XD-Violence:

These results further corroborate our method's robust generalization across diverse abnormal behaviors, including both prominent categories like Riot and visual concealed behaviors such as Abuse.

Effectiveness in Multi-event Scenarios: Please refer to Q3.

Robustness to Single-frame Annotation Bias: Please refer to Q2.

Adaptability across Different Backbones: To further evaluate FPL's effectiveness with varied backbone architectures, we conduct an additional experiment utilizing a CNN-based backbone from CU-Net [4] in UCF-Crime. The results are presented below:

As the results indicate, FPL surpasses SOTA methods with both CNN-based and Transformer-based lightweight backbones, demonstrating ideal adaptability across different model architectures.

Cross-dataset Transferability: Due to the inconsistency of anomaly definitions, cross-dataset validation resembles an open-set problem, rather than evaluation of generalization. Nevertheless, to thoroughly address the reviewer's request, we conduct additional experiments:

The results indicate a commendable cross-dataset transferability. This is primarily because single-frame supervision provides a clean and precise learning target. Instead, the learning process of previous MIL-based methods is vulnerable to shifted domains.

Response to Q5: Difference from Previous 2-stage MIL Methods

Previously, weakly-supervised VAD methods [4, 5] employed a 2-stage training strategy with pseudo labels, attempting to alleviate impact of noise. However, such methods inevitably converge to erroneous anomaly patterns with groundless top-k selection due to the deficiency of coarse-grained weak labels. In contrast, FPL first introduces evidential learning to estimate the anomaly relevance to the annotated precise abnormal patterns. Subsequently, intact abnormal intervals are mined based on the dynamic anomaly relevance and static feature similarity to reveal an accurate abnormal learning target. This resolves the issue of indistinct classification boundary caused by top-k selection, resulting in higher AUC and lower false alarms.

Response to Q6: Lightweight Architecture Design In fact, our model is concise with effective temporal modeling as in Sec. F in Supp. The model parameter size (in million) and inference time on the UCF-Crime test set are listed as follows:

Remarkably, FPL enables a lightweight network with a smaller parameter size and much faster inference time, to capture more accurate anomaly cues.

[1] Zhang, Huaxin, et al. "Glancevad: Exploring glance supervision for label-efficient video anomaly detection." arXiv, 2024.

[2] Liu, Kun, and Huadong Ma. "Exploring background-bias for anomaly detection in surveillance videos." ACM MM, 2019.

[3] Cui, Ran, et al. "Video moment retrieval from text queries via single frame annotation." SIGIR, 2022.

[4] Zhang, Chen, et al. "Exploiting completeness and uncertainty of pseudo labels for weakly supervised video anomaly detection." CVPR, 2023.

[5] Li, Shuo, et al. "Self-training multi-sequence learning with transformer for weakly supervised video anomaly detection." AAAI, 2022.

Thanks to the authors for the additional experiments and detailed explanations addressing my concerns. Unfortunately, the supplemental experimental results seem to further confirm my worries: under the single-frame annotation setting, the core of the model essentially becomes retrieving segments most similar to the annotated anomalous frame within a video, which fundamentally reduces the VAD task to a frame retrieval problem. This approach contradicts the fundamental goal of video anomaly detection, which is to achieve generalization and inductive reasoning across diverse scenarios, and thus diminishes the method’s practical value in real-world settings involving unknown or complex anomalies.

We sincerely appreciate the reviewer's thorough and insightful assessment, particularly the recognition of the effectiveness of SF-VAD. To respond the constructive questions, we'd like to offer the following clarifications：

Response to Q1: Choice of Thresholds

In Abnormal Event Mining, we utilize thresholds $\mathbf{\theta}$ to control the key frame selection process by frame-wise feature similarity. First, $\theta_1$ serves as a critical threshold to evaluate the discriminative capability of the feature similarity for identifying distinct abnormal intervals. A low variance suggests a consistent scene and motion pattern throughout the video, implying a reduced likelihood of multiple distinct abnormal intervals and thus limiting the reliability of multi-interval probing. In such instances, the system adaptively shifts its focus to precisely mine similar frames in the vicinity of the single annotated abnormal frame. The hyperparameter analysis of $\theta_1$​ is shown below:

Second, $\theta_2$ controls the degree of similarity required. A greater value of $\theta_2$ indicates that only frames exhibiting a higher degree of similarity are considered as alike abnormal clips.

Third, $\theta_3$ controls the gap of the relative position of selected key frames, which delegates the prominent key frames within each abnormal interval. Considering both computational overhead and mining precision, the key frame selection process is designed to sample only the most crucial frame within each abnormal interval for subsequent interval mining. Consequently, the gap between selected key frames is managed.

As the results show, the setting of thresholds $\theta$ achieve an ideal trade-off between interval mining completeness and accuracy, and a preferred generalization capability over multi-interval anomalies as Fig. 5 shows.

Response to Q2: Comparison with Fully-supervised Paradigm

The reviewer gains a nuanced insight and points out a rather interesting topic in VAD fields that fully-supervised paradigm doesn't perform ideally. Noticing such results are also observed in other literature [1, 2]. The key reasons lie in following aspects.

First, the annotation of precise abnormal intervals faces fundamental dilemmas, including inconsistent boundary annotation, omission of multi-interval abnormal events. In the one hand, the definition of abnormal event boundary is ambiguous. For instance, the interweaving of flame and smoke preceding and after the explosion anomaly are hard to categorize. The ambiguity leads to imprecise and inconsistent annotation, as mentioned in lines 28-31. It hinders the learning of a coherent decision boundary. On the other hand, an abnormal video may contain multiple abnormal events, and previous annotations [3, 4] often face the omission of anomalies due to the heavy burden of fine-grained annotation. This problem is particularly pronounced in fully supervised learning, where models tend to fit strictly to the provided labels. Consequently, some anomalies might be incorrectly classified as normal, leading to severe convergence issues during training.

Second, the fully-supervised paradigm faces severe generalizability issues. As abnormal behaviors can vary greatly within a domain (e.g., Explosion,  Riot in UCF-Crime), fully-supervised approaches impose strict supervision during training, which often leads to problems like overfitting. Moreover, the limited scale of fully-supervised VAD datasets due to high annotation costs exacerbates this issue. Details of the dataset statistics can be found in Sec. C in Supp.

Third, VAD datasets face severe data imbalance. The duration of abnormal videos and normal videos in the train set is shown below:

Since abnormal behaviors are inherently difficult to collect, the vast majority of frames in VAD datasets are normal. Given that fully-supervised models fit all labels, this imbalance causes them to exhibit a strong bias towards predicting frames as normal.

Due to the above issues, the experimental results and application prospects of the fully-supervised paradigm are unfortunately not ideal. Additional ablations are conducted:

In Contrast, SF-VAD fundamentally circumvents these issues by precisely annotating only a single abnormal frame, thereby reducing annotation cost, ensuring accuracy, and mitigating annotation bias. Through its exploration of anomaly pattern learning, it achieves  generalizability in multiple aspects and by inherently balancing normal and abnormal samples in its loss function, it demonstrates promising capabilities across annotation cost, detection performance, generalizability, and real-world applications.

Response to Q3: Annotation Efficiency

In fact, the addition of annotation cost of SF-VAD compared to weakly-supervised VAD is minor. As described in Sec.B in the Supp., the annotators are allowed to freely navigate in the video to locate a random abnormal frame. As people must witness an abnormal frame first, before they can make sure that the video is abnormal, the time addition is negligible. Dataset statistic in Sec. 4.2 is aligned with the protocol where the annotated abnormal frames is located at the front of the abnormal interval mostly.

Response to Q4: Method's Robustness

To address the reviewer's concerns, we evaluate the robustness of our method from the following aspects. Robustness to Single-Frame Annotation Bias: we conduct additional experiments to validate our method's robustness under varying single-frame annotation qualities. We sampled abnormal frames from GlanceVAD [6] and ARG [3] in UCF-Crime:

The results show that FPL's performance remains highly consistent, even with randomly selected ambiguous or non-representative single-frame annotations. Unlike previous methods [6, 7] that rely on a fixed distribution, my approach is significantly more position-agnostic because it dynamically deduces abnormal intervals. This design intrinsically enhances robustness against annotation variability.

Cross-Dataset Transferability: While cross-dataset validation often presents an open-set problem due to inconsistent anomaly definitions, we perform additional experiments to rigorously address the reviewer's query regarding single-frame anomaly generalization:

Given the fundamental difference in anomaly definitions between ShanghaiTech and UCF-Crime (where behaviors like biking, skateboarding, and chasing are considered abnormal in ShanghaiTech but normal in UCF-Crime), the first cross-dataset result shows a notable decrease. However, in all other cross-dataset evaluations, results demonstrate a commendable cross-dataset transferability for FPL. This superior generalization capability stems from the proposed single-frame supervision, which provides a clean and precise learning target. In contrast, the learning processes of previous MIL-based methods are often vulnerable to domain shifts, highlighting FPL's advantage in adapting to unseen anomalous events.

Response to Q5: Effectiveness with Different Backbones

To further evaluate the effectiveness of FPL using different backbones, we add additional experiments using CNN-base network from CU-Net [5], the results are shown below.

As the result shows, FPL surpasses SOTA methods with a lightweight CNN backbone. The superiority of performance mostly comes from the accurate learning target enabled by the FPL, instead of complex backbone design.

[1] Samaila, Yau Alhaji, et al. "Video anomaly detection: A systematic review of issues and prospects." Neurocomputing, 2024.

[2] Mostafa, Iman, et al. "Abnormal Human Activity Recognition in Video Surveillance: A Survey." Port-Said Engineering Research Journal, 2024.

[3] Liu, Kun, and Huadong Ma. "Exploring background-bias for anomaly detection in surveillance videos." ACM MM, 2019.

[4] Landi, Federico, Cees GM Snoek, and Rita Cucchiara. "Anomaly locality in video surveillance." arXiv, 2019.

[5] Zhang, Chen, et al. "Exploiting completeness and uncertainty of pseudo labels for weakly supervised video anomaly detection." CVPR, 2023.

[6] Zhang, Huaxin, et al. "Glancevad: Exploring glance supervision for label-efficient video anomaly detection." arXiv, 2024.

[7] Cui, Ran, et al. "Video moment retrieval from text queries via single frame annotation." SIGIR, 2022.

We're sincerely grateful to the reviewer for their thorough assessment. Especially, we appreciate the affirmation of the innovative SF-VAD paradigm that addresses the practical challenges in VAD and the effective FPL framework. Here is our response to the reviewer's productive comment.

Response to Q1: Clarification on Dataset Configuration

We select the three most challenging VAD datasets, ShanghaiTech, UCF-Crime, and XD-Violence, that are well-recognized by the VAD community to validate the effectiveness of SF-VAD on a large-scale. The videos remain identical to those of the original datasets, with the single abnormal frame annotation being extended.

Data Splits: we followed the original configurations for UCF-Crime and XD-Violence. For ShanghaiTech Campus, which was originally a semi-supervised dataset without abnormal videos in its training split, we adopted the widely applied data split created by Zhong et al. [1].

Dataset Annotation: we adopt an authentic annotation protocol to thoroughly assess SF-VAD's effectiveness under practical conditions. Under this protocol, annotators were permitted to freely navigate within abnormal event occurrences to pinpoint the single abnormal frame. This approach maximized annotation efficiency in applied settings while inherently reflecting annotation bias. Further details can be found in Sec. B in Supp.

Data Preprocessing: we adhere to a well-established paradigm. Initially, video frames are batched into 16-frame stacks. Subsequently, these frames are resized such that their shortest edge is 256 pixels, followed by 224-pixel crop augmentations. The processed video stacks are then input to a feature backbone for appearance or motion feature extraction. To manage sequence length variations, zero-padding or truncation is applied while ensuring the inclusion of the annotated abnormal frame.

Response to Q2: Discussion on Method Limitations

The discussion of limitations is provided in Sec. M of Supp. While our SF-VAD paradigm demonstrates robustness under varied scenarios, we acknowledge several promising avenues for future enhancement. The current multi-interval anomaly exploration primarily relies on individual video frame features; thus, an opportunity exists to further leverage inter-video correlations between abnormal annotations for richer pattern discovery. Additionally, integrating direct action traits could further enhance detection robustness, particularly for subtle anomalies. Due to the restrictions of the rebuttal policy, we are unable to provide specific failure case studies at this moment. However, we promise to include these in the revision.

Response to Q3: Code Organization and Documentation

Currently, a README.md file is included in the Supp. to provide an initial overview. Due to the rebuttal policy, we are unable to update code repository at this stage. However, upon publication, we are fully committed to releasing a user-friendly codebase, including the well-structured source code, detailed environment requirements and installation instructions, and all relevant annotation files that allow the VAD community to further investigate the potential of SF-VAD.

Response to Q4: Generalization of Various Anomaly Types

To validate the generalizability of different anomaly types, we demonstrate performance across different types of anomalies in XD-Violence, as follows:

For further evaluation, we performed additional experiments for the class-wise AUC metric on UCF-Crime. The results are depicted as below:

Notably, our method surpasses state-of-the-art methods in a variety of abnormal categories. This includes subtle abnormal behaviors like Arrest and Assault, which involve no dramatic scene changes, as well as more pronounced events such as Explosion or Arson, where scenes typically involve dramatic changes and clear visual cues. Under both circumstances, our proposed SF-VAD paradigm and FPL method demonstrate compelling performance, which reaffirms the overall generalizability of the proposed methods. In contrast to previous methods [2, 3], which might adapt techniques like Gaussian splatting around the annotated frame, our approach ensures that FPL can adapt to various abnormal behaviors and dynamically explore the anomaly duration. We achieve this by adopting evidential learning to dynamically measure the relevance to the annotated abnormal behaviors and integrating frame-wise feature similarity accordingly to mine the most probable abnormal duration.

Response to Q5: More Complete Evaluation

We appreciate the reviewer's valuable suggestion regarding additional evaluation. We contend that our current selection of three widely-used, real-world datasets, UCF-Crime, ShanghaiTech, and XD-Violence, was deliberate and has been widely used and thoroughly validated by previous VAD methods for comprehensive evaluation. These datasets are robust, encompassing diverse domains ranging from campus surveillance to real-world crime scenes and cinematic productions. They feature a broad spectrum of anomaly behaviors, include scenarios with significant interference, and cover perspectives that vary from static to highly dynamic. This comprehensive coverage enables a thorough assessment of our method's performance across various challenging conditions. More details can be found in Sec. E in Supp.

Authentic Annotation Protocol: Our annotation process is designed to reflect genuine human interests in a realistic setting. The principle of our annotation protocol is to examine the effectiveness of SF-VAD in practical scenarios, where annotators are allowed to freely navigate the video timeline and locate a random abnormal frame as soon as they witness an anomaly, without any preference guidance. Such a principle ensures that the annotation evades prior guidance that could artificially inflate performance. The detailed description of the annotation process and requirements has been provided in Sec. 4.1 and Sec. B in Supp., alongside dataset statistics in Sec. 4.2 and Sec. C in Supp.

Cross-Dataset Transferability: To further validate SF-VAD's generalizability and practical feasibility, we specifically examined its cross-dataset transferability.

The results indicate a commendable cross-dataset transferability. This is primarily because single-frame supervision provides an unambiguous definition of the anomaly, ensuring a clean and precise learning target that is less susceptible to noise and irrelevant contextual information, which aids in generalizing across varied datasets.

Computation Efficiency: Furthermore, our evaluation included assessing the method's computational efficiency.

Remarkably, our SF-VAD method demonstrates a smaller parameter size, much faster inference time, and better detection performance than previous state-of-the-art semi-supervised and weakly-supervised VAD methods. This superior performance-cost trade-off stems from the precise abnormal learning target enabled by FPL, which integrates anomaly relevance estimation with feature similarity to deduce more accurate learning targets than the top-k selection in MIL, which is less robust to noise interference. As a result, it allows a lightweight network architecture, without computationally expensive memory modules or text encoders, to effectively capture diverse anomalous cues.

[1] Zhong, Jia-Xing, et al. "Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection." CVPR, 2019.

[2] Zhang, Huaxin, et al. "Glancevad: Exploring glance supervision for label-efficient video anomaly detection." arXiv, 2024.

[3] Cui, Ran, et al. "Video moment retrieval from text queries via single frame annotation." SIGIR, 2022.

Dear Reviewers, Could you please review the authors’ rebuttal and let them know if you have any further comments or concerns?Do you feel your original comments have been adequately addressed? If not, it would be helpful to highlight any remaining issues. I would be grateful if you could kindly engage in the discussion to help move the review process forward. Best regards, AC