Deep Learning for Computer Vision

Thinking about localization as a regression problem

One fundamental way to think about localization is modeling the problem as a regression problem. The bounding box is four numbers and hence can be predicted in a direct manner with a setting for regression. We will also need to predict the label, which is a classification problem.

There are different parameterizations available to define the bounding boxes. There are four numbers usually for the bounding box. One of the representations is the center of the coordinates with the height and width of the bounding box. A pre-trained model can be used by removing the fully connected layer and replacing it with a regression encoder. The regression has to be regularized with the L2 loss which performs ...