Skip to content

Latest commit

 

History

History
20 lines (12 loc) · 1.39 KB

README.md

File metadata and controls

20 lines (12 loc) · 1.39 KB

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.

alt text

Parameters

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