CoReVAD: A Contextual Reasoning Framework for Training-Free Video Anomaly Detection

Official PyTorch implementation of "CoReVAD: A Contextual Reasoning Framework for Training-Free Video Anomaly Detection" (ICPR 2026) by Hyeongmuk Lim and Youngbum Hur.

Abstract: Existing Video Anomaly Detection (VAD) methods typically rely on task-specific training, leading to strong domain dependency and high training costs. Moreover, most existing methods output only scalar anomaly scores, providing limited insight into why specific events are considered abnormal.Recent advances in Vision– Language Models (VLMs) have enabled both anomaly detection and human-interpretable reasoning. However, many VLM-based approaches still require additional training steps (e.g., instruction tuning or verbalized learning) or external Large Language Models (LLMs), incurring further training costs and inference overhead. To address these challenges, we propose CoReVAD, a contextual reasoning framework for training-free video anomaly detection that operates with a single frozen VLM. CoReVAD directly generates anomaly scores and temporal descriptions from the VLM. To mitigate noise in generative outputs, we introduce a Local Response Cleaning (LRC) module based on local vision– text alignment. Furthermore, global temporal context and progression are incorporated through softmax-based refinement, Gaussian smoothing, and position weighting. Experiments on UCF-Crime and XD-Violence demonstrate that CoReVAD achieves competitive performance among training-free methods while providing reliable and interpretable explanations.

Data

For datasets, Please download the original videos from links (GT of each datasets is already included).

• UCF-Crime: link
• XD-Violence: link

The test video directory structure is as follows:

UCF-Crime
    └── videos
          ├── Abuse028_x264.mp4
          ├── Abuse030_x264.mp4
          └── ...
XD-Violence
    └── videos
          ├── A.Beautiful.Mind.2001__00-25-20_00-29-20_label_A.mp4
          ├── A.Beautiful.Mind.2001__00-40-52_00-42-01_label_A.mp4
          └── ...

Setup

Clone the repo

git clone https://github.com/Muk-00/CoReVAD.git
cd CoReVAD
conda create --name CoReVAD python=3.9
conda activate CoReVAD
pip install -r requirements.txt

Install the environment

In this paper, we use InternVL2, we follow the official installation instructions provided by InternVL2 (link)

Extract CLIP features

We first CLIP vision features from the dataset.

For UCF-Crime:
cd src/ucf
python extract_clip_features.py

For XD-Violence:
cd src/xd
python extract_clip_features.py

Output (UCF-Crime):

src/ucf/CLIP_feats
    └── ucf_test
          ├── Abuse028_x264_CLIP_features.npy
          ├── Abuse030_x264_CLIP_features.npy
          └── ...

Output (XD-Violence):

src/xd/CLIP_feats
    └── xd_test
          ├── A.Beautiful.Mind.2001__00-25-20_00-29-20_label_A_CLIP_features.npy
          ├── A.Beautiful.Mind.2001__00-40-52_00-42-01_label_A_CLIP_features.npy
          └── ...

Inference

1. Generate VLM response

The VLM responses are obtained in JSON format.

python generate_VLM_response.py

2. Local Response Cleaning (LRC)

The results of Local Response Cleaning (LRC) are saved in JSON format. We provide the generated responses in VLM_responses_LRC_ucf.json for UCF-Crime and VLM_responses_LRC_xd.json for XD-Violence.

python LRC.py

3. Evaluation

Evaluation for UCF-Crime dataset

eval_ucf.py

Evaluation for XD-Violence dataset

eval_xd.py

Acknowledgements

Thanks to Ye et al. for sharing their code.