In the "slam for dummies" tutorial, laser scanner was used, and two methods of landmark extraction were shown.
But most practical SLAM implementations are based on camera images. In these applications how are landmarks extracted? The Durrant-Whyte paper does not mention it and I could not find an end-to-end demo project that includes landmark extraction.
Usually landmarks are many distinct, salient features like blobs or corners within an image/frame. Salient features in computer vision are an art in itself. There are dozens of different approaches to extract, describe and match such features (you obviously need to recognize them in different frames).
Wikipedia has an overview and here you see a video of a practical SLAM implementation running on a mobile phone (a little bit of advertising for an excellent university:).
In case of images, they uses feature matching algorithms like SURF or ORB. ORB SLAM, one of the most popular open source slam solution uses ORB feature extraction only.
In vision-based applications, firstly features (interest points) are detected. The feature description step detects salient features only such as corners, blobs etc. and it doesn't make sense alone. Feature descriptors identify each feature in the scene. From this moment, you can choose the ones you want as landmark, keeping in mind your SLAM algorithm.