<<<<<<< HEAD FCN (Fully Convolutional Network for Semantic Segmentation) can classify images at the pixel level and solve the problem of image segmentation at the semantic level, so most of the existing semantic segmentation methods are based on FCN. However, these methods also have certain drawbacks, such as low resolution, lack of contextual information, and boundary errors. In 2020, related scholars constructively proposed OCRNet, a network framework based on Object Contextual Representation (OCR), in order to solve the problem of lack of contextual information in semantic segmentation. Its overall structure is shown below. Implementing this OCR method needs to go through three stages—first forming Soft Object Regions, then calculating Object Region Representations, and finally obtaining object contextual feature representation and contextual information-enhanced feature representation (Augmented Representation) . Compared with other semantic segmentation methods, the OCR method is more efficient and accurate. Because the OCR method solves the object region classification problem, not the pixel classification problem, that is, the OCR method can effectively and explicitly enhance the object information. In terms of performance and complexity, OCRNet is also better. In 2020, the "HRNet + OCR + SegFix" version won the first place in 2020ECCV Cityscapes.
For details, please refer toObject-Contextual Representations for SemanticSegmentation.
FCN（Fully Convolutional Network for Semantic Segmentation）可以对图像进行像素级的分类，解决了语义级别的图像分割问题，因此现有的大多数语义分割方法都基于FCN。但这些方法也有一定缺陷，比如分辨率低、上下文信息缺失和边界错误等。2020年，相关学者为解决语义分割上下文信息缺失难题，建设性地提出OCRNet，即基于物体上下文特征表示（Object Contextual Representation，以下简称OCR）的网络框架。其整体结构如下所示。实现此OCR方法需要经历三个阶段——首先形成软物体区域（Soft Object Regions），然后计算物体区域表示（Object Region Representations），最后得到物体上下文特征表示和上下文信息增强的特征表示（Augmented Representation）。 与其他语义分割方法相比，OCR方法更加高效准确。因为OCR方法解决的是物体区域分类问题，而非像素分类问题，即OCR方法可以有效地、显式地增强物体信息。从性能和复杂度来说，OCRNet也更为优秀。2020年，“HRNet + OCR + SegFix”版本在2020ECCV Cityscapes 获得了第一名。
具体原理细节请参考Object-Contextual Representations for SemanticSegmentation。
9c8570af (add new models)