A compositional approach to learning part-based models of objects

We propose a new method for learning probabilistic part-based models of objects using only a limited number of positive examples. The parts correspond to HOG bundles, which are groupings of HOG features. Each part model is supplemented by an appearance model, which captures the global appearance of the object by using bags of words of PHOW features. The learning is invariant to scaling and in-plane rotations of the object, the number of parts is learnt automatically, and multiple models can be learnt to allow for variations of 3D viewpoint or appearance. Through an experiment, we show that 3D multi-view object recognition can be performed by a series of learnt 2D models. The method is supervised but can learn models for multiple object viewpoints without these viewpoints being labeled in the training data. We evaluate our method on three benchmark datasets: (i) the ETHZ shape dataset, (ii) the INRIA horse dataset, and (iii) a multiple viewpoint car dataset. Our results on these datasets show proof of concept for our approach since they are superior or close to the state-of-the-art on all three datasets while we do not use any negative examples.