EgoCom Dataset v1.0

The EgoCom Dataset as described in the T-PAMI paper entitled "EgoCom: A Multi-person Multi-modal Egocentric Communications Dataset" available at https://ieeexplore.ieee.org/document/9200754.

This release contains five variations of the EgoCom dataset -- the only difference is the resolution of the videos to support varying download speed and storage capabilities:

• egocom1080p_uncompressed (total size: 267 GB)
  • Uncompressed raw 1080p (1920x1080 MP4 videos
• egocom720p (total size: 86 GB)
  • Compressed 720p (1280x720) MP4 videos
• egocom480p (total size: 28 GB)
  • Compressed 480p (640x480) MP4 videos
• egocom240p (total size: 9.5 GB)
  • Compressed 240p (352x240) MP4 videos
• egocom_audio_only (total size: 17 GB)
  • Original raw unnormalized, 16-bit, 44.1 kHz, steroeo WAV audio
  • Useful if you need audio only or if your spatial AI / conversational AI task needs the original magnitude of each audio signal

We also release the dataset of pre-trained embedding features used in the EgoCom Paper:

• egocom_pretrained_features (total size: 47 GB)
  • For every 1 second of each video in EgoCom, we provide text, video, and audio pre-trained feature embeddings.
  • These feature embeddings can be used to reproduce the results in the EgoCom Paper. See Section 4.2 for details on how they are computed.

Download the EgoCom Dataset

Make sure pigz and wget are installed:

# on Mac OS
brew install wget pigz
# on Ubuntu
sudo apt-get install pigz

To download and decompress the EgoCom dataset use the egocom_dataset/download_egocom.sh script:

./download_egocom.sh DATASET /path/to/egocom/

where DATASET is one of egocom1080p_uncompressed, egocom720p, egocom480p, egocom240p, egocom_pretrained_features, egocom_audio_only

Example usage:

./download_egocom.sh egocom720p .