2016/09/01 by Fatemehsadat Saleh, Mohammad Sadegh Ali Akbarian, Saleh, Fatemehsadat +9 · 4 citations
Computer Science · #Advanced Image and Video Retrieval Techniques #Advanced Neural Network Applications #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Visual Attention and Saliency Detection
paper · pdf · doi:10.48550/arxiv.1609.00446
openalex publication_date 2016/09/01 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Pixel-level annotations are expensive and time consuming to obtain. Hence,\nweak supervision using only image tags could have a significant impact in\nsemantic segmentation. Recently, CNN-based methods have proposed to fine-tune\npre-trained networks using image tags. Without additional information, this\nleads to poor localization accuracy. This problem, however, was alleviated by\nmaking use of objectness priors to generate foreground/background masks.\nUnfortunately these priors either require training pixel-level\nannotations/bounding boxes, or still yield inaccurate object boundaries. Here,\nwe propose a novel method to extract markedly more accurate masks from the\npre-trained network itself, forgoing external objectness modules. This is\naccomplished using the activations of the higher-level convolutional layers,\nsmoothed by a dense CRF. We demonstrate that our method, based on these masks\nand a weakly-supervised loss, outperforms the state-of-the-art tag-based\nweakly-supervised semantic segmentation techniques. Furthermore, we introduce a\nnew form of inexpensive weak supervision yielding an additional accuracy boost.\n