2019/11/28 by Theodora Kontogianni, Kontogianni, Theodora, Michael Gygli +5 · 2 citations
Computer Science · #Advanced Neural Network Applications #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Visual Attention and Saliency Detection
paper · pdf · doi:10.48550/arxiv.1911.12709
openalex publication_date 2019/11/28 · openalex created_date 2022/07/26 · openalex updated_date 2026/07/28
In interactive object segmentation a user collaborates with a computer vision\nmodel to segment an object. Recent works employ convolutional neural networks\nfor this task: Given an image and a set of corrections made by the user as\ninput, they output a segmentation mask. These approaches achieve strong\nperformance by training on large datasets but they keep the model parameters\nunchanged at test time. Instead, we recognize that user corrections can serve\nas sparse training examples and we propose a method that capitalizes on that\nidea to update the model parameters on-the-fly to the data at hand. Our\napproach enables the adaptation to a particular object and its background, to\ndistributions shifts in a test set, to specific object classes, and even to\nlarge domain changes, where the imaging modality changes between training and\ntesting. We perform extensive experiments on 8 diverse datasets and show:\nCompared to a model with frozen parameters, our method reduces the required\ncorrections (i) by 9%-30% when distribution shifts are small between training\nand testing; (ii) by 12%-44% when specializing to a specific class; (iii) and\nby 60% and 77% when we completely change domain between training and testing.\n