Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CADE: Continual Weakly-supervised Video Anomaly Detection with Ensembles

Satoshi Hashimoto, Tatsuya Konishi, Tomoya Kaichi, Kazunori Matsumoto, Mori Kurokawa

TL;DR

CADE tackles continual weakly-supervised video anomaly detection by integrating domain incremental learning with WVAD through a Dual-Generator and a Multi-Discriminator ensemble, enhanced by an inference-time ensemble to mitigate forgetting. The approach addresses data imbalance, label uncertainty, and incompleteness under domain shifts, and demonstrates strong performance on ShanghaiTech, CHAD, and UCF-Crime compared to baseline WVAD methods and approaching joint-training upper bounds. Ablation studies confirm the contributions of the DG and MD components, as well as the effectiveness of replay and ensemble strategies. The work advances practical WVAD deployment in changing environments while preserving privacy by avoiding real past-data storage.

Abstract

Video anomaly detection (VAD) has long been studied as a crucial problem in public security and crime prevention. In recent years, weakly-supervised VAD (WVAD) have attracted considerable attention due to their easy annotation process and promising research results. While existing WVAD methods tackle mainly on static datasets, the possibility that the domain of data can vary has been neglected. To adapt such domain-shift, the continual learning (CL) perspective is required because otherwise additional training only with new coming data could easily cause performance degradation for previous data, i.e., forgetting. Therefore, we propose a brand-new approach, called Continual Anomaly Detection with Ensembles (CADE) that is the first work combining CL and WVAD viewpoints. Specifically, CADE uses the Dual-Generator(DG) to address data imbalance and label uncertainty in WVAD. We also found that forgetting exacerbates the "incompleteness'' where the model becomes biased towards certain anomaly modes, leading to missed detections of various anomalies. To address this, we propose to ensemble Multi-Discriminator (MD) that capture missed anomalies in past scenes due to forgetting, using multiple models. Extensive experiments show that CADE significantly outperforms existing VAD methods on the common multi-scene VAD datasets, such as ShanghaiTech and Charlotte Anomaly datasets.

CADE: Continual Weakly-supervised Video Anomaly Detection with Ensembles

TL;DR

CADE tackles continual weakly-supervised video anomaly detection by integrating domain incremental learning with WVAD through a Dual-Generator and a Multi-Discriminator ensemble, enhanced by an inference-time ensemble to mitigate forgetting. The approach addresses data imbalance, label uncertainty, and incompleteness under domain shifts, and demonstrates strong performance on ShanghaiTech, CHAD, and UCF-Crime compared to baseline WVAD methods and approaching joint-training upper bounds. Ablation studies confirm the contributions of the DG and MD components, as well as the effectiveness of replay and ensemble strategies. The work advances practical WVAD deployment in changing environments while preserving privacy by avoiding real past-data storage.

Abstract

Video anomaly detection (VAD) has long been studied as a crucial problem in public security and crime prevention. In recent years, weakly-supervised VAD (WVAD) have attracted considerable attention due to their easy annotation process and promising research results. While existing WVAD methods tackle mainly on static datasets, the possibility that the domain of data can vary has been neglected. To adapt such domain-shift, the continual learning (CL) perspective is required because otherwise additional training only with new coming data could easily cause performance degradation for previous data, i.e., forgetting. Therefore, we propose a brand-new approach, called Continual Anomaly Detection with Ensembles (CADE) that is the first work combining CL and WVAD viewpoints. Specifically, CADE uses the Dual-Generator(DG) to address data imbalance and label uncertainty in WVAD. We also found that forgetting exacerbates the "incompleteness'' where the model becomes biased towards certain anomaly modes, leading to missed detections of various anomalies. To address this, we propose to ensemble Multi-Discriminator (MD) that capture missed anomalies in past scenes due to forgetting, using multiple models. Extensive experiments show that CADE significantly outperforms existing VAD methods on the common multi-scene VAD datasets, such as ShanghaiTech and Charlotte Anomaly datasets.

Paper Structure

This paper contains 16 sections, 8 equations, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Video Anomaly Detection
  4. Generative Replay
  5. Ensemble
  6. Proposed Method
  7. Formulation of MIL in DIL
  8. Dual-Generator
  9. Multi-Discriminator
  10. Ensemble
  11. Experiments
  12. Results on ShanghaiTech (SHT)
  13. Results on Charlotte Anomaly Dataset (CHAD)
  14. Results on UCF-Crime Dataset
  15. Ablation Studies
  16. ...and 1 more sections

Figures (4)

  • Figure 1: Our proposed method CADE is robust to additional training and maintains detection accuracy. While existing UR-DMU causes forgetting as its score becomes inactive when learning multiple scenes, CADE successfully indicates anomalies.
  • Figure 2: Our CADE consists of three key components: Dual-Generator, Multi-Discriminator, and an inference-time ensemble. Following the DIL setup, the model continuously learns using data divided by scene (domain). CADE can be easily integrated with existing WVAD methods by incorporating Dual-Generator and paired discriminators.
  • Figure 3: SHT results
  • Figure 4: CHAD results