The Development of Deep Learning Technologies by Center for Electronics & Information Studies Chinese Academy of Engineering

Author:Center for Electronics & Information Studies, Chinese Academy of Engineering
Language: eng
Format: epub
ISBN: 9789811545849
Publisher: Springer Singapore

3.Semantic segmentation. Different from image-level or object-level classification in image classification or target detection, semantic segmentation attempts to provide a pixel-based annotation that is a dense prediction for an image. Industrial applications thrive in unmanned autonomous vehicles and augmented reality. Most methods stem from a fully convolutional network (FCN) [30]. An FCN is an end-to-end model that directly calculates the probability of each pixel. Improvements such as U-Net and dilated convolution were later proposed. In recent years, a series of FCNs, including DeepLab [31] by Google and PSPNet [32] by CUHK/SenseTime, have become popular. A recent trend in this field is to optimize the results and to accelerate the prediction. In addition to semantic segmentation, which ignores the difference of each instance coming from the same semantic concept, instance segmentation has become popular to differentiate one instance from another. Mask R-CNN [33], created by Kaiming He, is a milestone method that appends a segmentation branch to faster R-CNN. In general, semantic segmentation is suitable for an amorphous region (denoted as “stuff”) segmentation such as the sky, while instance segmentation solves the countable object (denoted as “thing”) segmentation. A new task solving “stuff” and “thing” simultaneously, named Panoptic segmentation [34], has occurred very recently and attracted attention among Chinese researchers [35, 36].

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.