*Important.* This dataset is part of the [**torchmil** library](http://torchmil.readthedocs.io/).


Traffic Anomaly Detection (TAD) dataset adapted for Multiple Instance Learning (MIL).


### About the Original Traffic Anomaly Detection (TAD) Dataset

The original [Traffic Anomaly Detection (TAD) dataset](https://www.kaggle.com/datasets/nikanvasei/traffic-anomaly-dataset-tad) contains video clips. Each clip is labeled to indicate whether it contains an anomaly or not; however, frame-level labels are not available 

### Dataset Description

We have preprocessed the videos by computing features for each frame using various feature extractors.

- A **video** is labeled as positive (`label=1`) if it contains evidence of traffic anomaly.
- A **video** is labeled as positive (`label=1`) if it contains at least one positive frame.

This means a video is considered positive if there is any evidence of traffic anomaly.

### Directory structure

The following directory structure is expected:

```
root
├── features
│   ├── features_{features}
│   │   ├── video1.npy
│   │   ├── video2.npy
│   │   └── ...
├── labels
│   ├── video1.npy
│   ├── video2.npy
│   └── ...
└── splits.csv
```

Each `.npy` file corresponds to a video. The `splits.csv` file defines train/test splits for standardized experimentation.