Research map / Video / Video segmentation and tracking

UCF-Crime-DVS: A Novel Event-Based Dataset for Video Anomaly Detection with Spiking Neural Networks

AAAI 2025 · Video › Video segmentation and tracking · approach: Weakly supervised video anomaly detection · 11 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Builds UCF-Crime-DVS, the first event-camera version of UCF-Crime (1900 event streams, 13 anomaly classes), and proposes MSF, a fully spiking multi-scale network for weakly supervised video anomaly detection.

Problem

Video anomaly detection has relied on RGB and optical flow features; event cameras (DVS) are unexplored for this task, and ANNs handle asynchronous event streams poorly.

Approach

Re-records UCF-Crime by playing it on a monitor to an event camera, then uses a fully spiking network (multi-scale dilated spiking convolutions, a spiking GCN for global temporal dependencies, and a Temporal Interaction Module) with multiple-instance learning to score anomalies.

Key results

• MSF achieves 65.01% AUC and 3.27% FAR on UCF-Crime-DVS, versus 62.78% AUC and 11.52% FAR for the SNN baseline.
• Best SNN-based result among compared methods; ANN baselines such as AR-Net reach 60.71% AUC with 8.51% FAR.
• Ablation: adding TIM to LSF raises AUC from 62.44% to 64.39% and lowers FAR from 5.06% to 2.80%.

Measured effects reported in the paper

Surprising findings

• Adding TIM alone reduces AUC to 50.06% while giving the lowest FAR (0.68%), showing an AUC/FAR trade-off.
• Deeper SNN architectures did not improve VAD performance, suggesting network complexity alone does not help.

Limitations

• Performance remains well below state-of-the-art RGB methods on UCF-Crime, as the authors acknowledge.
• Events from near-static segments are not triggered, causing partial loss of anomaly information.
• Dataset is recorded by replaying the RGB video on a monitor, so it is not captured natively from real scenes.

Experimental setup

Datasets: UCF-Crime-DVS HARDVS

Benchmarks: UCF-Crime-DVS

Metrics: AUC FAR

Baselines: Sultani et al. 3C-Net AR-Net Wu et al. RTFM TSA SEW-ResNet PLIF SNN baseline

Builds on: Spikingformer Temporal Interaction Module dynamic multiple-instance learning leaky integrate-and-fire neuron SpikingJelly

Closest papers by meaning

• SFOD: Spiking Fusion Object Detector (CVPR 2024 · similarity 0.87)
• Hybrid Spiking Vision Transformer for Object Detection with Event Cameras (ICML 2025 · similarity 0.87)
• Efficient Event-Based Object Detection: A Hybrid Neural Network with Spatial and Temporal Attention (CVPR 2025 · similarity 0.87)
• MoniTor: Exploiting Large Language Models with Instruction for Online Video Anomaly Detection (NeurIPS 2025 · similarity 0.85)
• Advancing Video Anomaly Detection: A Concise Review and a New Dataset (NeurIPS 2024 · similarity 0.85)
• CREST: An Efficient Conjointly-trained Spike-driven Framework for Event-based Object Detection Exploiting Spatiotemporal Dynamics (AAAI 2025 · similarity 0.85)