CS 281B Project Report: Marginalized Kernels for Object Recognition

It is relatively easy to obtain sets of images that are known to contain objects of a certain category, such as “face”, or “bicycle”. There has been promising work on using such data sets — without any manual labeling of object features — to learn generative models for object categories [4, 1]. These models can then be used to detect instances of the categories in other images. However, these generative models may not be the best tools for discriminating between, say, bicycles and motorcycles. In this report, we investigate a discriminative approach that builds on the generative models. Specifically, we use a marginalized kernel that sums over possible matchings of image regions to object parts. We find that the marginalized kernel performs much better than the generative model alone on distinguishing between object categories, improving accuracy by up to 14 percentage points.