CNN based model of object detection for BDD100K, the source code support many configuration.
We used mobilenet_v2 as backbone, but change its stride(==2) of first convolution layer into stride==1. The whole net is similar to RefineDet, but we don't classify the background and object in refinement module, we just want the anchors to catch as many as objects in refinement module, and then use the up-sample to build the object detection module.
