---
language:
  - en
license: apache-2.0  
tags:
  - video-anomaly-detection
  - computer-vision
  - temporal-action-localization
  - ucf-crime
  - i3d-features
datasets:
  - sultani/real-world-anomaly-detection-in-surveillance-videos  
task_categories:
  - video-classification
size_categories:
  - 1GB<n 
pretty_name: UCF-Crime I3D Features
pipeline_tag: video-classification  
widget:
  - example_title: Example Video Features
    example_code: |
      import h5py
      import numpy as np
      with h5py.File('ucf_crime_features_labeled.h5', 'r') as f:
          features = np.array(f['Abuse/Abuse001_x264/features'][:10])  
          labels = np.array(f['Abuse/Abuse001_x264/labels'][:10])
      print("Features shape:", features.shape)
      print("Labels:", labels)
subdir_list:
  - abuse
  - arrest
  - assault
  - arrest_features
  - arson_features
  - assault_features
  - burglary_features
  - explosion_features
  - fighting_features 
  - normal_features
  - roadaccidents_features
  - robbery_features
  - shooting_features
  - shoplifting_features
  - stealing_features
  - vandalism_features
  
data_files:
  - train: ucf_crime_features_labeled.h5  # Single file with splits via metadata
  - validation: ucf_crime_features_labeled.h5
  - test: ucf_crime_features_labeled.h5
---



# UCF-Crime: Precomputed I3D Features with Temporal Annotations

This dataset provides pre-extracted 1024-dimensional I3D RGB features along with frame-level temporal anomaly labels for videos from the UCF-Crime dataset.

---

## Dataset Characteristics

### **Features**
- 1024-dimensional I3D RGB feature vectors  
- Extracted from **64 uniformly sampled frames per video**  
- Feature tensor shape: **[64, 1024]** 

### **Temporal Annotations**
- Mapped from original anomaly intervals  
- Re-scaled to match the 64 sampled frames  
- Only videos with valid annotations are included  

### **Coverage**
- Videos that contain complete temporal anomaly intervals  
- Suitable for supervised learning tasks

---

## Recommended Usage

This dataset is ideal for:

- Frame-level binary classification  
- Reconstruction-based anomaly detection  
- Temporal convolutional networks (TCN)  
- Transformer-based sequence models  
- Sequential anomaly scoring models  

Since features are already extracted, experiments are **lightweight and GPU-efficient**.

---

## Loading the Dataset
The Data Loader code has also been provided. Please refer to that.


---
## Citation
@inproceedings{sultani2018real,
  title={Real-world Anomaly Detection in Surveillance Videos},
  author={Sultani, Waqas and Chen, Chen and Shah, Mubarak},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={4469--4478},
  year={2018}
}