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 |
