SmaCS is currently involved in the final round of recordings in the cabin mock-ups, with an updated recording protocol, based on the conclusions obtained from the evaluation of the preliminary version, with special emphasis on the lessons learned from the recording sessions of October 2020. Participants will simulate the set of situations that might happen within the cabin, in which the luggage is in correct and incorrect positions. Several takes will be recorded with a high variety of people and luggage appearances, under different illumination conditions and points of view.
The data collection in SmaCS is performed by combining three kind of data sources: The first source is a publicly available annotated datasets to train initial versions of the descriptor and to help to understand the details for optimal annotation of content of interest from the images. The second source are the recordings in the cabin mock-ups. Finally, the third source are synthetic images provided by the visualization tool developed in the project. Real and synthetic images will be combined by means of a synthetic-to-real domain adaptation procedure to train the system.
This enhanced version of the data obtained from the cabin recordings will contain more classes (kind of objects), more samples per class, improved annotations (i.e., better suited for the training stage), higher and more balanced appearance, pose and illumination variations. Its scale is expected to be at least in the order of hundreds of thousands of samples.