We propose an approach to reduce the runtime of detection methods such as Exemplar-SVM that apply an extensive number of detectors to the same image.

One can view a query image as a set of candidate windows, for all of which the detector responses need to be computed.

While each candidate window is most likely unique, due to the high patch recurrence apparent in images there could be many shared parts between these candidate windows.

We leverage this redundancy and avoid duplicate calculations for similar window parts. 

Instead, all similar parts are assigned the same detector response and the overall response for each window is approximated through the shared responses.

Experiments on the PASCAL VOC07 detection task show that our approximation maintains the accuracy of Exemplar-SVM while reducing runtime significantly.