The idea of outputting the coordinates of points of interest in the image using CNNs can be extended to many other applications. Some of these include human-pose estimation (DeepPose: Human Pose Estimation via Deep Neural Networks), as shown:

The keypoints/landmarks are defined for the objects in the training images. These keypoint locations must be consistent for a particular object among all training images.

For example, in facial-keypoint detection, say we are interested in locating the eyes, nose, and mouth, we must define a number of key points around the eyes, nose, and mouth of all the training face images. We then train the CNN to output the predicted keypoint locations, just like in the preceding image, and then a regression loss is applied to these output key point coordinates to train the CNN. At test time, the input image is fed into CNN to predict all the keypoint positions. The following illustration shows facial-keypoint detection: