Vehicle Detection Onboard Small Unmanned Aircraft - Advances in Embedded Computer Vision - page 199

Fig. 9.4 This shows one of the rotated test images as provided to the detector trained with the
“aligned” training set
Figure 9.4 shows the rotated image as it is applied to the aligned detector. This
enlarges the size of the image, increases the number of search windows, and increases
the computational cost of vehicle detection by adding the step of several image
rotations before applying the detector repeatedly.
9.3.3 Rotation Invariance Through Rotated Training Set
The secondmethod attempts to learn a detector that is rotationally invariant so that the
detector has to be applied only once, irrespective of vehicle orientation. The rotated
training set consists of multiple transformations of each image in the aligned training
set: each image is rotated through 360 ◦ in 5 ◦ increments. To maintain a consistent
size for all rotational increments during detector training, all images are the same
square dimensions.
Figure 9.5 shows an example of the rotation performed for one of the training set
images. This increased the size of the positive training image set from 200 images
to 14,400 images.