Vitto Ferrari
Thu 25 Jun 2015, 12:45 - 14:00

If you have a question about this talk, please contact: Steph Smith (ssmith32)

Object detection with the Ensemble of Exemplar-SVMs: optimal calibration and contextual guidance

The Ensemble of Exemplar-SVMs (EE-SVM) is a powerful non-parametric approach to object detection. It is widely used because it explicitly associates a training example to each object it detects in a test image. This enables transferring meta-data such as segmentation masks, 3D models, and viewpoint labels. In the first part of the talk we present a technique for calibrating the EE-SVM model. While the standard approach calibrates each SVM independently, our method optimizes their joint performance as an ensemble. We formulate joint calibration as a constrained optimization problem and devise an efficient optimization algorithm to find its global optimum. Experiments on the ILSVRC 2014 dataset shows that (i) our joint calibration procedure outperforms independent calibration on the task of classifying windows as belonging to an object class or not; and (ii) this better window classifier leads to better performance on the object detection task. As they have a separate component for each positive training sample, EE-SVMs trained on large datasets are slow at test time. In the second part of the talk, we present a context-driven technique to automatically select a subset of components most relevant for a given test image, based only on its global appearance. This leads to a 10x speed-up when running large models. Based on the same technique, we also estimate at which positions and scales the object is likely to appear and use it to improve detection performance.