Using Human Knowledge to Judge Part Goodness: Interactive Part Selection

It is a common practice to model object detectors as collection of filters. For these detectors to be effective, it is important to select “good filters” covering most of the variation of the data. In order to achieve this, these methods invest majority of their time selecting a good subset of filters from a large pool. Good filters are less cluttered and their gradients are spatially correlated. Humans can differentiate between a good filter and a bad one by visualizing them. In addition humans bring with them the knowledge of diversity and effectiveness of filters, properties which are difficult to model. In this work, we show that humans can help build better detectors by including their knowledge of good filters. We show this by building an interactive framework for poselet selection. Our interactive framework improves the detection performance on the PASCAL VOC dataset and significantly improves the training time.