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.
var | val |
---|---|
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 |
mAP | ~0.42 |