The task is to detect and clssify the objects present in the aerial images by determining their bounding boxes.

I took VGG16 with batch normalization as backbone model, and further trained the last few FC layers for classification.

Devide the image to 7x7(or 8x8) grids, each cell predicts two bounding-boxes and object confidence.

I use aerial images in DOTA dataset for object detection. And there are total 16 classes in this implementation.

OBJECT_CLASSES = {
    "plane": 0,
    "ship": 1,
    "storage-tank": 2,
    "baseball-diamond": 3,
    "tennis-court": 4,
    "basketball-court": 5,
    "ground-track-field": 6,
    "harbor": 7,
    "bridge": 8,
    "small-vehicle": 9,
    "large-vehicle": 10,
    "helicopter": 11,
    "roundabout": 12,
    "soccer-ball-field": 13,
    "swimming-pool": 14,
    "container-crane": 15
}

The model outputs in different training stages (i.e. early, middle, final stages) are as below:

• Tennis-court
• Plane
• Storage-tank