Boundary Detection Using F-Measure-, Filter- and Feature- (F3) Boost

In this work we propose a boosting-based approach to boundary detection that advances the current state-of-the-art. To achieve this we introduce the following novel ideas: (a) we use a training criterion that approximates the F-measure of the classifier, instead of the exponential loss that is commonly used in boosting. We optimize this criterion using Anyboost. (b) We deal with the ambiguous information about orientation of the boundary in the annotation by treating it as a hidden variable, and train our classifier using Multiple-Instance Learning. (c) We adapt the Filterboost approach of [1] to leverage information from the whole training set to train our classifier, instead of using a fixed subset of points. (d) We extract discriminative features from appearance descriptors that are computed densely over the image. We demonstrate the performance of our approach on the Berkeley Segmentation Benchmark.