We have developed a novel shape descriptor that has similar characteristics as Histogram of Oriented Gradients (HOG descriptor) but instead of densely sampling all shapes and structures we use only shapes that are relevant for properly describing visual categories. We find those relevant shapes by using a learning process of LHOP model which allows us to generate compositional shapes based on statistic of edges found in a general set of images. Our descriptors is then based on those learnt compositional shapes hence the name Histogram of Compositions.
We have used LHOP model to train compositions in hierarchical manner and produced vocabulary of compositions for 3 layers. The training images used are simple images with relatively clear edges which are required to properly learn compositions. All training images are also general enough to produce a set of compositions that can be used across multitude of different categories. This can be achieved due to small size of compositions in first three layers where they can easily be shared among different kind of categories. Examples of training images and produced vocabulary of compositions:
Based on learnt vocabulary the descriptor is then constructed using a simple region partitioning and histograms across one or more layers are calculated for each region:
Performance
We tested descriptor on Caltech-101 dataset and compared it against HOG descriptor:
