Can Giraffes Become Birds? An Evaluation of Image-to-image Translation for Data Generation
Daniel Ruíz, Gabriel Salomon, Eduardo Todt · Anais do XI Computer on the Beach - COTB '20 · 2020
There is an increasing interest in image-to-image translation withapplications ranging from generating maps from satellite images tocreating entire clothes’ images from only contours. In the presentwork, we investigate image-to-image translation using GenerativeAdversarial Networks (GANs) for generating new data, taking as acase study the morphing of giraffes images into bird images. Morphinga giraffe into a bird is a challenging task, as they have differentscales, textures, and morphology. An unsupervised cross-domaintranslator entitled InstaGAN was trained on giraffes and birds,along with their respective masks, to learn translation betweenboth domains. A dataset of synthetic bird images was generatedusing translation from originally giraffe images while preservingthe original spatial arrangement and background. It is important tostress that the generated birds do not exist, being only the result of alatent representation learned by InstaGAN. Two subsets of commonliterature datasets were used for training the GAN and generatingthe translated images: COCO and Caltech-UCSD Birds 200-2011.To evaluate the realness and quality of the generated images andmasks, qualitative and quantitative analyses were made. For thequantitative analysis, a pre-trained Mask R-CNN was used for thedetection and segmentation of birds on Pascal VOC, Caltech-UCSDBirds 200-2011, and our new dataset entitled FakeSet. The generateddataset achieved detection and segmentation results close tothe real datasets, suggesting that the generated images are realisticenough to be detected and segmented by a state-of-the-art deepneural network.