EfficientDet is better than Yolo v3, it has achieved 52.2% accuracy on the COCO dataset for the object detection tasks.
EfficientDet is a family of models which are more efficient, faster, and accurate in object detection.
Architecture:
It uses EfficientNet-B3 as the backbone and bi-directional feature pyramid network (BiFPN).
BiFPN
BiFPN takes input as P3, P4, P5, P6, P7. these Pn are features of the input image, the formula for finding the size of Pn is (Images Size / 2^n)
P3 = 1024/2^3 = 128
P4 = 1024/2^4 = 64
P5 = 1024/2^5 = 32
P6 = 1024/2^6 = 16
P7 = 1024/2^7 = 8
These features are resized first then aggregates, resize is either done by upsampling or downsampling.
Aggregation of these features is not like normal summation but it is weighted summation, and weights are learnable parameters. this method is called Fast Normalized Fusion, it is 1.26 to 1.31 times faster than softmax, still, give the same accuracy.
