2020/03/16 by Moritz Kampelmühler, Kampelmühler, Moritz, Axel Pinz +1 · 1 citation
Computer Science · #Advanced Image and Video Retrieval Techniques #Advanced Vision and Imaging #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Multimodal Machine Learning Applications #Visual Attention and Saliency Detection
paper · pdf · doi:10.48550/arxiv.2003.07101
openalex publication_date 2020/03/16 · openalex created_date 2022/07/26 · openalex updated_date 2026/07/28
Humans are able to precisely communicate diverse concepts by employing\nsketches, a highly reduced and abstract shape based representation of visual\ncontent. We propose, for the first time, a fully convolutional end-to-end\narchitecture that is able to synthesize human-like sketches of objects in\nnatural images with potentially cluttered background. To enable an architecture\nto learn this highly abstract mapping, we employ the following key components:\n(1) a fully convolutional encoder-decoder structure, (2) a perceptual\nsimilarity loss function operating in an abstract feature space and (3)\nconditioning of the decoder on the label of the object that shall be sketched.\nGiven the combination of these architectural concepts, we can train our\nstructure in an end-to-end supervised fashion on a collection of sketch-image\npairs. The generated sketches of our architecture can be classified with 85.6%\nTop-5 accuracy and we verify their visual quality via a user study. We find\nthat deep features as a perceptual similarity metric enable image translation\nwith large domain gaps and our findings further show that convolutional neural\nnetworks trained on image classification tasks implicitly learn to encode shape\ninformation. Code is available under\nhttps://github.com/kampelmuehler/synthesizinghumanlikesketches\n