Tesla neural network model hydranet

Backbone： Refers to the feature extraction network , Used to identify multiple objects in a single image , And provide rich characteristic information of the object . We use it a lot AlexNet、ResNet、VGGNet As a backbone network .
Detection Head(head)： In feature extraction （ The backbone ） after , It provides us with the input characteristic graph representation . For some practical tasks , For example, the detection object 、 Segmentation, etc . We usually apply a “ Detection head ”, So it's like a head attached to the trunk .
Neck：Neck Between the trunk and the head , It is used to extract some finer features .（ For example, feature pyramid network （FPN）,BiFPN）

Target detection has a general structure ：​​Input → backbone → neck → head → Output​​.

In Tesla neural network architecture ：

• backbone: RegNet + ResNet
• neck: BiFPN
• head: HydraNet

         Even though AutoML and NAS It works , But they also have limitations ：1) High resource consumption ,2) Poor flexibility ,3) Poor generalization ,4) The design result is difficult to understand .

Tesla Using a residual neural network block designed Regnet（ Regular network structure ） As the backbone of its neural network .

RegNet yes 2020 year Facebook AI research (FAIR) The paper Designing Network Design Spaces A new network design paradigm proposed in .
 

HydraNets There are three main benefits :

• Feature sharing ： Reduce the repeated convolution calculation , Reduce the number of trunks , Especially when testing, it is more efficient ;