AVSET-10M Dataset
Overview
The AVSET-10M dataset is a comprehensive collection of audio-visual samples designed for research in multimedia content analysis, audio-visual recognition, and machine learning. It is divided into two distinct subsets: AVSET-700K and AVSET-10M (excluding AVSET-700K). This dataset provides a rich set of meta-information, enhancing its utility for diverse research applications.
Dataset Composition
AVSET-10M is released as two subsets:
1. AVSET-700K
- Description: This subset consists of 727,530 audio-visual corresponding samples meticulously filtered from the AudioSet.
- Features:
- Each video segment is accompanied by a manually labeled audio category.
- Ensures accurate categorization and relevance of the audio-visual samples.
2. AVSET-10M (w/o. AVSET-700K)
- Description: This subset includes 9,877,475 audio-visual corresponding samples, filtered from the expansive Panda-70M dataset.
- Features:
- Focus on semantically coherent video segments that concentrate on a single event.
- Includes text descriptions sourced from the original Panda70M dataset.
- Pseudo-labels for audio categories are provided, derived using PANNs, along with their corresponding confidence scores.
AVSET-10M Dataset Structure
Repository for the AVSET-10M dataset, containing two subsets. Below is the directory structure with links to significant components:
AVSET-10M/
│
├── [AVSET-700K/](#avset-700k)
│ ├── unbalanced_train_segments_part00.json
│ ├── ...
│ └── unbalanced_train_segments_part40.json
│
├── [AVSET-10M(excluding-700K)/]
│ ├── split_000.json
│ ├── ...
│ └── split_399.json
│
└──[ontology.json]
Meta-Information
Each video clip in both subsets of the AVSET-10M dataset includes comprehensive meta-information:
- Youtube ID: YoutubeID to get the download url.
- Start Time: Specific start times for each video clip.
- End Time: Specific end times for each video clip.
- AVC: Measures the similarity between the audio and visual components of the clip.
- Flag (Sound Separation): Indicates whether sound separation is required for the clip. "1" for the need for sound separation.
- Label: Relevant textual descriptions or labels associated with each clip.
Additional Details for AVSET-10M (w/o. AVSET-700K)
- pseudo-label: Audio categories with confidence scores to assist in dataset partitioning and analysis.
License
See license. The video samples are collected from a publicly available dataset. Users must follow the related license to use these video samples.