Pixels are the basis of most modern image representations and the starting point of most computer vision algorithms. However, they are actually just artefacts of the image capture and display process. They do not correspond to entities in the real world. Enter superpixels. Superpixels are groupings of pixels into homogeneous regions. They are not yet semantic segmentations of images but reduce the complexity of an image by merging continuous regions into mid-level representations.
During my PhD I worked on some ideas around Superpixel segmentation based on shape-centered features. I’ve just added a summary of the ideas to the publications section. You can also download and read the poster.