Tag : #Deep Learning#Object Detection#Backbone
In short, the structure of the object detection model based on deep learning can be summarized as: input->backbone->neck->head->output
backbone network extracts features, neck extracts some more Complex features, then the head computes the predicted output
Backbone
We generally translate Backbone as a backbone network or backbone network , which mainly refers to the part of the network used for feature extraction. Generally, models that have been pre-trained on large datasets (such as ImageNet|COCO, etc.) are used, such as: ResNet-50, Darknet53, VGG, etc.
Neck
It is used between Backbone and Detection Head to better integrate and extract features, further improving the diversity and robustness of features. Some models add layers that collect feature maps at different stages.
Detection Head
- Get the output of the network and use the extracted features to make predictions
- The network layer of the subsequent connection mainly serves the detection task, so it is called " Detection head "
- The detection head is translated as the detection head is a network that obtains the output content of the model, and uses the previously extracted features to make predictions, mainly including the type and location (bounding boxes) of the predicted target .
other
Bottle Neck : Bottleneck, usually refers to the different dimensions of network input and output data, usually the output dimension is much smaller than the input dimension,
GAP : Global Average Pool global average pooling, average the features of a certain channel
Warm up : Small learning takes the lead in training Several epochs, this is because the parameters of the network are randomly initialized, and a large learning rate is used at the beginning to be prone to numerical instability