Based on SVIRO, we created additional images for three new vehicle interiors. For each vehicle, we randomly generated 250 training and 250 test scenes where each scenery was rendered under 10 different illumination and environmental conditions. We created two versions: one containing only people and a second one including additionally occupied child and infant seats. We used 10 different exterior environments (HDR images rotated randomly around the vehicles), 14 (or 8) human models, 6 (or 4) children and 3 babies respectively for the training and test split. The four infant and two child seats have the same geometry for each split, but they use different textures. Consequently, the models need to generalize to new illumination conditions, humans and textures. There are four possible classes for each seat position (empty, infant seat, child seat and adult) leading to a total of 4³=64 classes for the whole image.
The ground truth labelling is the same as for SVIRO. Each scenery has a dedicated folder with 10 variations of the same scene under different illumination conditions.
The following images show sceneries under 5 (out of 10) illumination variations per vehicle.
It is also possible to generate new ground truth data for already existing sceneries. Feel free to contact us with your ideas.