The first parts of this Thesis are focused on the study of the supervised evaluation of image segmentation algorithms. Supervised in the sense that the segmentation results are compared to a human-made annotation, known as ground truth, by means of different measures of similarity. The evaluation depends, therefore, on three main points.
First, the image segmentation techniques we evaluate. We review the state of the art in image segmentation, making an explicit difference between those techniques that provide a flat output, that is, a single clustering of the set of pixels into regions; and those that produce a hierarchical segmentation, that is, a tree-like structure that represents regions at different scales from the details to the whole image.
Second, ground-truth databases are of paramount importance in the evaluation. They can be divided into those annotated only at object level, that is, with marked sets of pixels that refer to objects that do not cover the whole image; or those with annotated full partitions, which provide a full clustering of all pixels in an image. Depending on the type of database, we say that the analysis is done from an object perspective or from a partition perspective.
Finally, the similarity measures used to compare the generated results to the ground truth are what will provide us with a quantitative tool to evaluate whether our results are good, and in which way they can be improved. The main contributions of the first parts of the thesis are in the field of the similarity measures.
First of all, from an object perspective, we review the used basic measures to compare two object representations and show that some of them are equivalent. In order to evaluate full partitions and hierarchies against an object, one needs to select which of their regions form the object to be assessed. We review and improve these techniques by means of a mathematical model of the problem. This analysis allows us to show that hierarchies can represent objects much better with much less number of regions than flat partitions.
From a partition perspective, the literature about evaluation measures is large and entangled. Our first contribution is to review, structure, and deduplicate the measures available. We provide a new measure that we show that improves previous ones in terms of a set of qualitative and quantitative meta-measures. We also extend the measures on flat partitions to cover hierarchical segmentations.
The second part of this Thesis moves from the evaluation of image segmentation to its application to object detection. In particular, we build on some of the conclusions extracted in the first part to generate segmented object candidates. Given a set of hierarchies, we build the pairs and triplets of regions, we learn to combine the set from each hierarchy, and we rank them using low-level and mid-level cues. We conduct an extensive experimental validation that show that our method outperforms the state of the art in many metrics tested.
