2019/02/27 by Jungbeom Lee, Eunji Kim, Lee, Jungbeom +7 · 3 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 #Machine Learning and Data Classification
paper · pdf · doi:10.48550/arxiv.1902.10421
openalex publication_date 2019/02/27 · openalex created_date 2022/07/29 · openalex updated_date 2026/07/28
The main obstacle to weakly supervised semantic image segmentation is the\ndifficulty of obtaining pixel-level information from coarse image-level\nannotations. Most methods based on image-level annotations use localization\nmaps obtained from the classifier, but these only focus on the small\ndiscriminative parts of objects and do not capture precise boundaries.\nFickleNet explores diverse combinations of locations on feature maps created by\ngeneric deep neural networks. It selects hidden units randomly and then uses\nthem to obtain activation scores for image classification. FickleNet implicitly\nlearns the coherence of each location in the feature maps, resulting in a\nlocalization map which identifies both discriminative and other parts of\nobjects. The ensemble effects are obtained from a single network by selecting\nrandom hidden unit pairs, which means that a variety of localization maps are\ngenerated from a single image. Our approach does not require any additional\ntraining steps and only adds a simple layer to a standard convolutional neural\nnetwork; nevertheless it outperforms recent comparable techniques on the Pascal\nVOC 2012 benchmark in both weakly and semi-supervised settings.\n