Bounding box detection of drones (small scale quadcopters)
Localization of quadcopters in images with CNTK's implemententation of Fast R-CNN. I collected a total of 500 images of the quadcopter DJI Phantom 3 from Google's image search and dozens of screenshots from YouTube videos. After that I manually added bounding boxes to the collected images. 350 images were used for training and 150 images for testing. A mAP of about 0.42 was achieved.
|number of regions of interest (cntk_nrRois)||500|
|minimum region of interest relative size (roi_minDimRel)||0.04|
|maximum region of interest relative size (roi_maxDimRel)||1.0 (had to increase this value)|
|number of images used for training||350|
|number of images used for testing||150|