Where2Stand: Toward a Framework for Portrait Position Recommendation in Photography

Composition layout is crucial in the portrait location recommendation of photography. The existing studies require both landscape background images and portrait foreground images, which limits the scope of practical applications. In this paper, we propose an end-to-end portrait location recommendation model, which mainly consists of three sub-networks: the first sub-networks is the portrait generation network, which generates relatively real portrait foreground images based on random input noise; the second sub-networks is the spatial transformation network, which mainly changes the size and location of the generated portrait based on the input landscape image; The third sub-networks is the compose network to generate a realistic portrait landscape image, which considers not only the correlation between the portrait foreground and the landscape background but also the overall composition aesthetics. Last, the proper standing position is obtained by computing the difference between the generated and input landscape images. We also construct a portrait landscape photo dataset PLDataset to train and verify our method. The experimental results on our dataset show that our proposed method can recommend a relatively reasonable standing position by only providing a landscape image in portrait landscape photography, which greatly increases the availability.