VGGSound-50k Preprocessed Dataset

This dataset contains preprocessed data from the VGGSound dataset, specifically processed using the VGGSound-AVEL50k subset for cross-modal knowledge distillation research. The preprocessing is optimized for MST-Distill (Mixture of Specialized Teachers for Cross-Modal Knowledge Distillation) method.

This preprocessing work is based on the VGGSound-AVEL50k subset from: jasongief/CPSP: [2023 TPAMI] Contrastive Positive Sample Propagation along the Audio-Visual Event Line

And related preprocessing works are described in our paper: MST-Distill: Mixture of Specialized Teachers for Cross-Modal Knowledge Distillation | Code

Original Dataset

The original VGGSound dataset and the AVEL50k subset are available at:

Original VGGSound: VGGSound Dataset
VGGSound-AVEL50k subset: Used in CPSP research for audio-visual event localization

Dataset Information

Classes: 141 audio-visual event categories
Samples: 48,755 video clips from VGGSound-AVEL50k subset (available at processing time)
Modalities: Audio and visual features
Content: Audio-visual events with temporal segment labels
Optimization: Specifically preprocessed for Cross-modal Knowledge Distillation (CMKD) methods

Preprocessing Details

The preprocessing pipeline consists of four main stages:

1. Data Organization (`preprocess_VGGsound50K_0.py`)

Maps video IDs to file paths and class labels
Creates category-to-index mapping for 141 classes
Output: VGGS50K_videos.txt with video paths and class labels

2. Feature Extraction Verification (`preprocess_VGGsound50K_1.py`)

Verifies availability of both audio and visual features
Ensures corresponding audio and visual features exist
Output: VGGS50k_metadata.txt with valid sample names and labels

3. Segment Label Processing (`preprocess_VGGsound50K_2.py`)

Converts segment labels from JSON format to numpy arrays
10-segment temporal labels for each video
Output: Individual .npy files for each sample's segment labels

4. Feature Normalization (`preprocess_VGGsound50K_3.py`)

Min-max normalization applied globally across all features
Normalizes both audio and visual features
Output: Normalized feature files in separate directories

Data Structure

File Structure

VGGS50K/
├── VGGS50K_features_normed/
│   ├── audio_features/
│   │   └── *_aFeature.npy          # Normalized audio features
│   └── visual_features/
│       └── *_vFeature.npy          # Normalized visual features
├── seg_labels/
│   └── *_sLabel.npy                # Temporal segment labels [1, 10]
├── VGGS50k_metadata.txt            # Sample names and class labels
└── VGGS50K_videos.txt              # Video paths and class labels

Usage

import numpy as np
import torch

# Load sample data
sample_name = "your_sample_name"
label = 0  # class label

# Load features
audio_features = np.load(f'VGGS50K_features_normed/audio_features/{sample_name}_aFeature.npy')
visual_features = np.load(f'VGGS50K_features_normed/visual_features/{sample_name}_vFeature.npy')
segment_labels = np.load(f'seg_labels/{sample_name}_sLabel.npy')

# Convert to PyTorch tensors
audio_tensor = torch.from_numpy(audio_features.astype(np.float32))
visual_tensor = torch.from_numpy(visual_features.astype(np.float32))

print(f"Audio features shape: {audio_tensor.shape}")
print(f"Visual features shape: {visual_tensor.shape}")
print(f"Segment labels shape: {segment_labels.shape}")

Applications

This preprocessed dataset is optimized for:

Audio-visual event localization with temporal segment labels
Cross-modal knowledge distillation

Features and Specifications

Audio Features: Normalized using global min-max scaling
Visual Features: Normalized using global min-max scaling
Temporal Resolution: 10-segment labels for event localization
Quality: Only samples with complete audio-visual features included

License

This preprocessed dataset maintains the same license as the original VGGSound dataset: Creative Commons Attribution 4.0 International License (CC BY 4.0).

Citation

If you use this preprocessed dataset in your research, please cite the original VGGSound paper and the CPSP paper:

Original VGGSound Citation:

@inproceedings{chen2020vggsound,
  title={Vggsound: A large-scale audio-visual dataset},
  author={Chen, Honglie and Xie, Weidi and Vedaldi, Andrea and Zisserman, Andrew},
  booktitle={ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={721--725},
  year={2020}
}

CPSP Paper Citation:

@article{zhou2022CPSP,
  title={Contrastive Positive Sample Propagation along the Audio-Visual Event Line},
  author={Zhou, Jinxing and Guo, Dan and Wang, Meng},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2022},
  publisher={IEEE}
}

Acknowledgments

We thank the original authors of the VGGSound dataset and the CPSP research team for making these valuable resources available to the research community.

Note to original authors: If you have any concerns or objections regarding this preprocessed dataset, please contact us and we will promptly remove it.