The Video Diver Detection dataset (VDD-C) is an enormous dataset comprised of approximately 105,000 fully annotated images of divers, drawn from videos taken in pool and field environments. This dataset improves on previous datasets in size (approximately 17 times more images), but also by providing images in the video context, allowing for analysis of the video-context stability of single-image detectors. Additionally, video data is important for robotic applications: photographer bias in photo-based datasets means that video datasets have more realistic translations, rotations of divers that can encourage the network to learn translation/rotation invariance.
A big thanks go to our collaborators who provided us access to the raw data, particularly to the McGill University Mobile Robotics Lab.
Check back soon for a link to the downloadable dataset!