COCO Dataset Stuff Segmentation Challenge

In computer vision, image segmentation is a method in which a digital image is divided/partitioned into multiple set of pixels which are called super-pixels, stuff segmentation challenge is a newly introduced task in which we have to segment out stuff out of the digital image. This job is to promote the state of the art in semantical category segmentation. Semantic segmentation defines the method by which each image pixel with the category is combined. Semantic classes can be well-defined stuff such as a vehicle, an individual or things like huge backdrop areas such as anarea covered with grass. stuff classes are crucial, as they explain aspects of an image such as type of scene and thing classes represent things, objects, there are various factors to it such as location, physical attributes, type of the material and properties of the scene. Only courses like lawn, ceiling, are focused on this problem. The method in this paper consists of a convolutional neural network and provides a superior framework pixel-level task, the dataset used in this research is the COCO dataset [1], which is used in a worldwide challenge on Codalab. It is a huge dataset with a total of 164k images in which around 118k images in the training dataset and around 5k images in the validation dataset, 20K images in test-dev , 20K images in test-challenge, we use this dataset to analyze the importance of thing and stuff classes in terms of their cover range and to check how many times they appear in image captions, this research is also important to find out whether stuff segmentation is easier than thing segmentation or not.