Predicting pixels

Image classification is the task of predicting labels or categories. Object detection is the task of predicting a list of several deep learning-based algorithms with its corresponding bounding box. The bounding box may have objects other than the detected object inside it. In some applications, labeling every pixel to a label is important rather than bounding box which may have multiple objects. Semantic segmentation is the task of predicting pixel-wise labels.

Here is an example of an image and its corresponding semantic segmentation:

  

>

As shown in the image, an input image is predicted with labels for every pixel. The labels could be the sky, tree, person, mountain, and bridge. Rather than assigning a label to the whole image, labels are assigned to each pixel. Semantic segmentation labels pixels independently. You will notice that every people is not distinguished. All the persons in the image are labeled in the same way.

Here is an example where every instance of the...