A PyTorch implementation of EfficientDet
A PyTorch implementation of EfficientDet architecture: EfficientDet: Scalable and Efficient Object Detection. Based on EfficientNet, newly introduce a weighted bi-directional feature pyramid network (BiFPN) and propose a compound scaling method that uniformly scales the resolution, depth, and width for all backbone, feature network, and box/class prediction networks at the same time. EfficientDet-D6 achieves state-of-the-art 50.9 mAP on COCO dataset with 52M parameters and 229B FLOPs, being 4x smaller and using 13x fewer FLOPs yet still more accurate (+0.2% mAP) than the best previous detector.
Backbone model EfficientNet is modified from my own implementation of EfficientNet-PyTorch. EfficientDet is based on the official implementation google/automl/efficientdet in Tensorflow.