The classification score is computed for every window in the sliding window approach or the fully convolutional approach to know what object is present in that window. Instead of predicting the classification score for every window to detect an object, each window itself can be predicted with a classification score. Combining all the ideas such as sliding window, scale-space, full convolution, and regression give superior results than any individual approach. The following are the top five localization error rates on the ImageNet dataset achieved by various networks using the regression approach:
The preceding graph shows that the deeper the network, the better the results. For AlexNet, localization methods were not described in the paper. The OverFeat used multi-scale convolutional regression with box merging. VGG used localization but with fewer scales and location. These gains are attributed to deep features. The ResNet uses a different localization method and much deeper features.
The regression encoder and classification encoder function independently. Hence there is a possibility of predicting an incorrect label for a bounding box. This problem can be overcome by attaching the regression encoder at different layers. This method could also be used for multiple objects hence solving the object detection problem. Given an image, find all instances in that. It's hard to treat detection as regression because the number of outputs are variable. One image may have two objects and another may have three or more. In the next section, we will see the algorithms dealing with detection problems more effectively.