Abstract

Since the invention of the Hough transform (HT) in 1960’s, this method has become a very popular method for object detection. Originally proposed for detecting parametric objects like lines, circles, etc., HT has been generalized for detecting arbitrary shapes by parameterizing the objects by a reference point, e.g. the center of mass, and accumulating the votes of local image patches for this point. Although this representation enforces consistency of patches in their relative locations, it produces false positives by combining votes that are consistent in location but inconsistent in other properties like shape, pose, color, etc. which is the main reason behind the poor performance of these approaches. This all raises several questions about the validity of the center as the only parametrization of object categories. In particular, is voting for the center necessary? What other criteria can be used for voting? Can we learn the optimal properties from data?

In this talk, I will propose the Latent Hough Transform (LHT) for enforcing consistency among votes. The idea behind LHT is to augment the Hough transform with latent variables and perform voting in a latent space instead of voting only for location. This way, only votes that agree on the assignment of the latent variables are allowed to support a hypothesis.  I will further show how to learn an optimal latent space from training data by exploiting the linearity of the Hough transform based methods. Our extensive experiments on challenging object detection benchmarks show that our proposed method outperforms traditional Hough transform based methods and even leads to state-of-the-art results on some categories.