Datasets:
Update README.md
Browse files
README.md
CHANGED
|
@@ -38,57 +38,12 @@ VCapAV addresses the lack of multimodal deepfake data involving environmental so
|
|
| 38 |
|
| 39 |
## Uses
|
| 40 |
|
| 41 |
-
VCapAV is intended for:
|
| 42 |
-
|
| 43 |
- Audio anti-spoofing research
|
| 44 |
- Audio-visual deepfake detection
|
| 45 |
- Evaluation of general-purpose audio generation methods
|
| 46 |
- Studying modality consistency between vision and sound
|
| 47 |
- Research on multimodal synchronization, scene-aware generation, and cross-modal alignment
|
| 48 |
|
| 49 |
-
## Dataset Structure
|
| 50 |
-
|
| 51 |
-
The VCapAV dataset consists of **three multimodal manipulation types**, each reflecting a different audio–visual deepfake scenario:
|
| 52 |
-
|
| 53 |
-
### **1. Fake Audio + Real Visual**
|
| 54 |
-
|
| 55 |
-
Environmental audio is manipulated while the original video remains unchanged.
|
| 56 |
-
|
| 57 |
-
Fake audio is generated using five state-of-the-art methods:
|
| 58 |
-
|
| 59 |
-
- **V2A-Mapper**
|
| 60 |
-
- **V2A-MLP**
|
| 61 |
-
- **AudioLDM1**
|
| 62 |
-
- **AudioLDM2**
|
| 63 |
-
- **AudioCraft (AudioGen)**
|
| 64 |
-
|
| 65 |
-
### **2. Real Audio + Fake Visual**
|
| 66 |
-
|
| 67 |
-
The visual stream is manipulated using the **Kling Large Text-to-Video model**, producing high-fidelity 1080p videos based on generated captions.
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
### **3. Fake Audio + Fake Visual**
|
| 71 |
-
|
| 72 |
-
Both audio and video are manipulated.
|
| 73 |
-
Each Kling-generated video is paired with all five fake-audio variants → **1,210 multimodal deepfake pairs**.
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
## Dataset Statistics
|
| 77 |
-
|
| 78 |
-
A detailed breakdown of the dataset composition:
|
| 79 |
-
|
| 80 |
-
| **Category** | **# Clips** | **Hours** |
|
| 81 |
-
|---------------------------|-------------|-----------|
|
| 82 |
-
| Real Video | 14,923 | 41.45 |
|
| 83 |
-
| Fake Video | 242 | 0.67 |
|
| 84 |
-
| Real Audio | 14,923 | 41.45 |
|
| 85 |
-
| Fake Audio | 74,615 | 207.26 |
|
| 86 |
-
| Fake Audio + Fake Video | 1,210 | 3.36 |
|
| 87 |
-
| **Total** | **90,990** | **252.75** |
|
| 88 |
-
|
| 89 |
-
All clips are **≤10 seconds**, audio is uniformly **downsampled to 16 kHz**, and the dataset spans **242 environmental sound classes**.
|
| 90 |
-
|
| 91 |
-
|
| 92 |
## Dataset Creation
|
| 93 |
|
| 94 |
Most deepfake datasets focus on speech or human faces. VCapAV fills this gap by focusing on **general environmental audio** and **video–audio consistency**, enabling research on non-speech deepfake detection.
|
|
|
|
| 38 |
|
| 39 |
## Uses
|
| 40 |
|
|
|
|
|
|
|
| 41 |
- Audio anti-spoofing research
|
| 42 |
- Audio-visual deepfake detection
|
| 43 |
- Evaluation of general-purpose audio generation methods
|
| 44 |
- Studying modality consistency between vision and sound
|
| 45 |
- Research on multimodal synchronization, scene-aware generation, and cross-modal alignment
|
| 46 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 47 |
## Dataset Creation
|
| 48 |
|
| 49 |
Most deepfake datasets focus on speech or human faces. VCapAV fills this gap by focusing on **general environmental audio** and **video–audio consistency**, enabling research on non-speech deepfake detection.
|