Keypoints are the important points of an image, and finding a keypoint representation of an image can be an important way of summarizing it. However, most unsupervised methods that produce a small number of keypoints (sparse representation) are generally unable to adapt to unseen types of images, in comparison to methods that produce dense representations. Thus, this project tries to construct a generic network that produces sparse keypoint representations, by using a pre-trained dense keypoint network's output and selecting a subset of the dense keypoint set. 

The network currently has been tested and works with simple images, and the next step is to test it on more complicated real-world and computer-generated images. If succesful, this will be the first generic unsupervised sparse keypoint network, and will have many applications in robotics. In particular, it will enable vision-controlled robots to operate in previously-unseen environments and have the same behavior as that in the environments they're trained on.