GPI Seminar Series: Guillem Palou

Guillem Palou, Detection-Classification Framework and applications to Depth Ordering Evaluation
Wednesday June 19th, at 12:00, MERIT room D5-010

Abstract:
In classification problems it is normal to present results using confusion matrices, where the missclassification of objects is observed among different classes. In detection frameworks, there normally exists a tradeoff between precision and recall. When a system architecture consists of first a detection step followed by a binary classification, it is likely that the classification score also depends on the precision-recall operating point of the first step. For instance, if only confident detections are considered (low recall, high precision) it is likely to obtain a high classification score. On the contrary, if many low-confident  results are detected, (low precision, high recall) it is also likely a decrease in classification performance. To integrate both the detection and classification problem, we introduce the concept of a precision-recall-classification framework to evaluate the performance of a system. We apply the proposed measures to the evaluation of state of the art depth ordering systems and show a clearer view on systems performance.