Nowadays state of the art detection methods in remote sensing are widely inspired by computer vision successful algorithms. Detectors such as Support Vector Machines (SVMs) (trained on Histogram of Oriented Gradients - HOGs - for example) have shown prime results in object detection and are now extensively used in remote sensing. More recent methods like Deformable Part Models (DPMs) have also led to good results in remote sensing. This paper investigates the various concepts behind the DPMs and propose an extended DPM that handle multi-modal data. We show that DPMs can advantageously be used with descriptors that suit the sensor used to generate the data and propose an approach for combining DPMs trained over various sensors.