Anomaly AR-Net: Weakly Supervised Video Anomaly Detection via Center-Guided Discriminative Learning

October 22, 2025 · View on GitHub

Language: English | 中文

Abstract

This repository implements Anomaly AR-Net, a novel weakly supervised video anomaly detection framework presented at ICME 2020. Our approach leverages center-guided discriminative learning to effectively detect anomalies in video sequences using only video-level labels. The method addresses the challenge of temporal localization in weakly supervised settings by incorporating attention mechanisms and multi-scale feature learning.

🎯 Key Features

Weakly Supervised Learning: Requires only video-level labels for training
Center-Guided Discriminative Learning: Enhances feature discrimination between normal and abnormal patterns
Multi-scale Temporal Modeling: Captures temporal dependencies at different scales
Attention Mechanisms: Focuses on relevant temporal segments
Multiple Model Architectures: Supports various backbone networks for feature extraction

🏗️ Model Architecture

The framework consists of several key components:

Core Models:

Model_single: Basic linear classifier with dropout
Model_mean: Multi-scale convolutional layers with average pooling
Model_sequence: Sequential convolutional network with residual connections
Model_concatcate: Multi-scale feature concatenation
model_lstm: Bidirectional LSTM for temporal modeling
BaS_Net: Background Suppression Network with attention mechanisms

Key Components:

Filter Module: Generates attention weights for temporal segments
CAS Module: Class Activation Sequence for temporal localization
Multi-scale Convolution: Captures features at different temporal resolutions
Attention Mechanisms: Focuses on relevant video segments

📁 Project Structure

Anomaly_AR_Net_ICME_2020/
├── model.py                    # Neural network model architectures
├── options.py                  # Command line argument parser
├── main.py                     # Main entry point and training setup
├── train.py                    # Training loop and optimization
├── test.py                     # Model testing and evaluation
├── losses.py                   # Custom loss functions
├── utils.py                    # Utility functions and helpers
├── video_dataset_anomaly_balance_uni_sample.py  # Dataset loading and processing
├── environment.yaml            # Conda environment configuration
├── LICENSE                     # MIT License
├── README.md                   # English documentation
└── README_CN.md                # Chinese documentation

🔧 Code Components

Core Modules

Model Architecture (`model.py`)

Multiple model variants for different temporal modeling approaches
Filter Module: Temporal attention mechanism for foreground/background separation
CAS Module: Class Activation Sequence for temporal localization
Multi-scale convolution: Captures temporal patterns at different resolutions
LSTM integration: Bidirectional LSTM for sequence modeling
Comprehensive documentation with bilingual comments (English/Chinese)

Training Pipeline (`train.py`)

Weakly supervised learning with video-level labels
Center-guided discriminative learning for feature separation
Multi-instance learning framework
Loss optimization with various loss functions
Detailed training loop with logging and checkpointing

Data Processing (`video_dataset_anomaly_balance_uni_sample.py`)

Temporal sequence sampling with balanced normal/abnormal samples
Feature extraction from pre-computed I3D features
Sequence padding for variable-length videos
Multi-dataset support (ShanghaiTech, UCF-Crime, Avenue)
Memory-efficient loading with optional data dictionary

Loss Functions (`losses.py`)

Discriminative loss functions for weakly supervised learning
Center-guided learning to enhance feature discrimination
Temporal consistency for smooth predictions
K-Max Multiple Instance Learning (KMXMILL) loss implementation

Utility Functions (`utils.py`)

Feature processing with random extraction and perturbation
Attention masking for variable-length sequences
Visualization tools for anomaly score plotting
Data preprocessing and normalization utilities

Configuration Management (`options.py`)

Comprehensive argument parsing for all training/testing parameters
Hardware configuration (GPU selection, memory settings)
Dataset and feature specifications
Training hyperparameters and optimization settings

Key Features

Weak Supervision

Video-level labels only for training
Temporal localization from weak supervision
Multi-instance learning paradigm
Balanced sampling of normal and abnormal videos

Temporal Modeling

Multi-scale temporal convolution for different time resolutions
Attention mechanisms for temporal focus
Sequence modeling with LSTM networks
Background suppression for improved anomaly detection

Feature Processing

I3D feature extraction for spatio-temporal representation
Multi-modal support (RGB, Flow, Combined features)
Feature normalization and preprocessing
Variable-length sequence handling with padding

Code Quality

Comprehensive documentation with bilingual comments
Modular architecture for easy extension
Type hints and clear variable naming
Error handling and validation

🚀 Quick Start

Installation

# Clone repository
git clone https://github.com/wanboyang/Anomaly_AR_Net_ICME_2020.git
cd Anomaly_AR_Net_ICME_2020

# Create environment
conda env create -f environment.yaml
conda activate anomaly_icme

Data Preparation

Download I3D features from:
- Baidu Netdisk (password: u4k6)
- Google Drive
Extract the dataset:
```
tar -xvf dataset.tar
```
Update dataset path in configuration

Visual Feature Extraction

To extract visual features similar to this project, clone:

git clone https://github.com/wanboyang/anomaly_feature

Training

python main.py

The models and testing results will be saved in ./ckpt and ./results directories respectively.

📊 Performance

Our method achieves state-of-the-art performance on multiple video anomaly detection benchmarks:

UCF-Crime: Competitive performance in weakly supervised setting
ShanghaiTech: Effective anomaly localization
Avenue: Robust detection across different anomaly types

📚 Citation

If you find this work useful for your research, please cite:

@inproceedings{anomaly_wan2020icme,
  title={Weakly Supervised Video Anomaly Detection via Center-Guided Discriminative Learning},
  author={Wan, Boyang and Fang, Yuming and Xia, Xue and Mei, Jiajie},
  booktitle={Proceedings of the IEEE International Conference on Multimedia and Expo},
  year={2020}
}

🎥 Video Presentation

Watch the oral presentation on Bilibili

🤝 Acknowledgements

We thank the contributors of W-TALC and the PyTorch team for their excellent frameworks.

📧 Contact

For questions and suggestions, please contact:

Boyang Wan - wanboyangjerry@163.com