The EfficientDet architecture was written by Google Brain. EfficientDet s built on top of EfficientNet, a convolutional neural network that is pretrained on the ImageNet image database for classification. EfficientDet pools and mixes portions of the image at given granularities and forms features that are passed through a NAS-FPN feature fusion layer. The NAS-FPN combines various features at varying granularities and passes them forward to the detection head, where bounding boxes and class labels are predicted.
EfficientDet is a family of models expressing the same architecture at different model size scales. The paper carefully explores the tradeoffs in scaling and object detection model. Do you make the ConvNet deeper? The feature fusion neck wider? The image resolution higher? How do you balance all of these scaling factors in the most efficient manner?
We implement EfficientDet here within the TensorFlow 2 Object Detection API. The TensorFlow 2 Object Detection API allows you to quickly swap out different model architectures, including all of those in the EfficientDet model family and many more.
An EfficientDet model trained on the COCO dataset yielded results with higher performance as a function of FLOPS.
For a deeper dive see: https://blog.roboflow.ai/breaking-down-efficientdet/ and https://blog.roboflow.ai/the-tensorflow2-object-detection-library-is-here/