Bilinear image translation for temporal analysis of photo collections

History-lapse videos

Examples of frames from history-lapse videos generated by our bilinear adversarial translation model. Here, input images of cars from the 1940s are translated to different future times by varying the input time variable. The output is a video showing changes of car style over time. Please note how the shape and appearance of the input cars changes over time. For example, the curved hood gradually fattens, the round lights become rectangular, the windscreen gets bigger, the front grille gets smaller and the running board on the side gradually disappears. Note also that other characteristics of the input car instances, such as viewpoint, color and the overall shape are preserved.