On structuring multiple grouping hypotheses in generic object detection

A novel contour grouping method was recently proposed for the difficult task of detecting and delineating unexpected multi-part objects of unknown specific shape and appearance in a variety of natural images. This method, in many ways original and unique, was generally able to obtain object-level groups of quite good quality for a variety of objects and images. For each tested image, a number of object-level groups are hypothesized and ranked using a generic multicriteria objective function. Experiments shown that object-level groups most similar to the human ground truth usually rank high. However, no object-level group was obtained with some difficult images. This paper proposes three important improvements to that original method. Firstly, fixed parameters are replaced by adaptive parameters, improving the robustness of the method even for the most difficult images. Secondly, a parallel version of the method is developed to either speed-up or scale-up the computation, making the method adaptive to practical space and time constraints. Finally, a further structuring of the object-level groups makes it possible to isolate the interesting ones and determine their number. A comparison with previous results illustrates the significance of the improvements.