| # 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](https://github.com/qiuqiangkong/audioset_tagging_cnn), 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](LICENSE). The video samples are collected from a publicly available dataset. Users must follow the related [license](https://raw.githubusercontent.com/microsoft/XPretrain/main/hd-vila-100m/LICENSE) to use these video samples. | |