README.md

---
license: other
task_categories:
  - audio-classification
tags:
  - audio
  - sound-event-detection
  - environmental-sound
  - urban-sound
  - glass-breaking
  - gunshots
  - screams
size_categories:
  - 1K<n<10K
---

# MIVIA DB4 – Audio Surveillance Event Dataset

> ⚠️ **Private repository** — This dataset is derived from the MIVIA DB4 
> research dataset (University of Salerno). Not for public redistribution.

## Dataset Summary

MIVIA DB4 is a benchmark dataset for **acoustic surveillance event detection** 
in urban environments. It contains annotated scene recordings with three target 
event classes embedded in realistic background noise.

**Original dataset:** MIVIA Lab, University of Salerno  
**Reference:** Foggia et al., IEEE TIFS 2016

---

## Statistics

| | Train | Test | Total |
|---|---|---|---|
| Scenes | 66 | 29 | 95 |
| Events | 2,100 | 900 | 3,000 |
| Glass events | 700 | 300 | 1,000 |
| Gunshot events | 700 | 300 | 1,000 |
| Scream events | 700 | 300 | 1,000 |
| Audio format | 32000 Hz mono | 32000 Hz mono | — |
| Scene duration | ~180 s | ~180 s | — |

**Perfectly balanced** — 1,000 events per class, 70/30 train/test split.

## Event Classes

| CLASS_ID | Category | Description |
|---|---|---|
| 1 | background | Ambient background (not a target event) |
| 2 | glass | Glass breaking sounds |
| 3 | gunshots | Gunshot sounds |
| 4 | screams | Human screams |

## Background Subclasses

`bells`, `cars`, `crowd`, `crowd_claps`, `gaussian_noise`, 
`household_app`, `rain`, `twistle`

---

## Files

| File | Description |
|---|---|
| `data/train/*.wav` | 528 training scene recordings |
| `data/test/*.wav` | 232 testing scene recordings |
| `DB4_train_events.csv` | 2,100 training event annotations |
| `DB4_test_events.csv` | 900 testing event annotations |
| `DB4_all_events.csv` | All 3,000 events combined |
| `metadata.csv` | Scene metadata (powers Dataset Viewer) |

## CSV Columns

| Column | Example | Description |
|---|---|---|
| `xml_file` | `00001.xml` | Source annotation file |
| `split` | `train` | train or test |
| `scene_base` | `00001` | Scene ID |
| `event_pathname` | `glass/0001.wav` | Isolated event sound |
| `event_category` | `glass` | glass / gunshots / screams |
| `class_id` | `2` | Numeric class label |
| `start_sec` | `5.3647` | Event start time (s) |
| `end_sec` | `5.8497` | Event end time (s) |
| `bg_subclass` | `cars` | Background type |
| `bg_pathname` | `background/cars/0026.wav` | Background file |
| `audio_exists` | `true` | WAV found on disk |

---

## Usage in Colab

```python
import pandas as pd
from huggingface_hub import hf_hub_download

# Load training annotations
train_path = hf_hub_download(
    repo_id="Titung/MIVIA-GENERAL",
    filename="DB4_train_events.csv",
    repo_type="dataset"
)
train_df = pd.read_csv(train_path)
print(train_df["event_category"].value_counts())
```

```python
# Load a scene audio file and slice an event
import librosa
from huggingface_hub import hf_hub_download

wav_path = hf_hub_download(
    repo_id="Titung/MIVIA-GENERAL",
    filename="data/train/00001_00.wav",
    repo_type="dataset"
)
y, sr = librosa.load(wav_path, sr=None)

row = train_df.iloc[0]
event = y[int(row.start_sec * sr) : int(row.end_sec * sr)]
print(f"Class: {row.event_category} | Duration: {len(event)/sr:.3f}s")
```

```python
# Load with HF datasets library
from datasets import load_dataset

ds = load_dataset(
    "Titung/MIVIA-GENERAL",
    data_files={"train": "DB4_train_events.csv",
                 "test":  "DB4_test_events.csv"}
)
print(ds)
```

---

## Citation

> Foggia, P., Petkov, N., Saggese, A., Strisciuglio, N., & Vento, M. (2016).  
> *Recognizing and Localizing the Sounds of Abnormal Events in Urban Environments.*  
> IEEE Transactions on Information Forensics and Security, 11(5), 1026–1037.

MIVIA Lab, University of Salerno — http://mivia.unisa.it