HybridNets: 

End-to-End Perception Network

    End-to-end Network has become increasingly important in multi-tasking. One prominent example of this is the growing significance of a driving perception system in autonomous driving. 

    In this post we will briefly step into the big world of this kind of networks, called hybrid nets.

Introduction    

    Recent advances in embedded systems' computational power and neural networks' performance have made autonomous driving an active field in computer vision. Ideally, to create a vehicle capable of driving itself is to feed it with every bit of information available in its immediate surroundings.

Methodology 

    Let's have a look on the architecture of those end-to-end networks that can multitask.

As shown in the figure our one-stage network includes one sharing encoder and two separated decoders

to solve distinct tasks. 

Experimentation and evaluation

    The dataset for three tasks is prepared according to existing multi-task networks trained on BDD100K to aid in comparison. Of all the ten classes in object detection, only {car, truck, bus, train} is selected and merged into a single class {vehicle} since DLT-Net and MultiNet can only detect vehicles.

Now let's have a look on how efficient line detection is with hybrid nets.

Reference: 

https://paperswithcode.com/paper/hybridnets-end-to-end-perception-network-1

Some code to play around with:

https://github.com/datvuthanh/HybridNets