Shock! EfficientDet finally open up!

Official propaganda! Remember that EfficientDet Tu list do? Today it came! Numerous github up before the main attempt to reproduce but they stop at the reproduction Starter Edition D0, and today has been the official release of the source code!

 

EfficientDet results

As a model Tu chart, it naturally has its innovations, not to mention lightweight, CPU friendly terminal and friendly back-end network architecture is more refined series of feature extraction methods, such as BiFPN:

 

 

Let's see how EfficientDet results:

D6 which can achieve amazing 50.9, unprecedented in the coco list!

Even YoloV3, MaskRCNN, RetinaNet, NAS-FPN in front of it had to beats.

 

Official Code

The official release has been based tensorflow code. We can not only refer to the official of reproduction to achieve their own version, it can be expected there will be more top models in the near future based on efficientdet to develop!

Students fast download them! Github link:

google / automl github.comicon

Paper:

 

https://arxiv.org/pdf/1911.09070.pdf​arxiv.org

 

 

Posted on 2020-03-18

Published 482 original articles · won praise 789 · Views 1.71 million +

Guess you like

Origin blog.csdn.net/weixin_42137700/article/details/105266395