Description, analysis and evaluation of confidence estimation procedures for sub-categorisation

This report presents contributions in two main areas: the combination of low level image features with visual saliency maps and use of confidence measures for information fusion. For the first part, we show experiments with automatic saliency estimation methods based on bottom-up and top-down approaches. We also explored maps generated by mouse-clicks. These maps were used to give weights to local image features which are then used for image categorisation in an approach based on bag-of-patches. For the second part, we explored methods to associate confidence values to predictions of each test sample. These values are used to give per-sample weights for different information sources in classifier combination.