Improving object detection performance on hard to detect instances in DOTA

Authors

• Said Harb
• Tim Schmidt

Dataset

The dataset used in our experiments is the Dota Dataset. In particular we use the version 1.5 of this dataset. It consists of 2.806 Aerial images with different sizes. The image sizes vary from 800x800 to 20.000 x 20.000 pixels. On these images there are 16 classes labeled which are:

• large vehicle
• small vehicle
• helicopter
• plane
• ship
• swimmingpool
• container crane
• storage tank
• bridge
• harbor
• roundabout
• baseball-diamond
• basketball court
• ground track field
• tennis court
• soccerball field

Model(s)

The model employed is the YOLOv5n.

The newest version and information about the model can be found here: Link

Results

Both employed approaches improved the detection performance on hard to detect instances considerably. However the oversampling apporach prooved to be slightly better, as the metrics are the best for this experiment.

The following images show the mAP overall and the AP per class for the best experiment: