YOLO v2 SSD should be and the same period of the paper, compared YOLO v1 and RCNN series, SSD draws on the advantages of both, and joined the multi-scale to compensate for yolo in the detection of small objects lack in accuracy and speed are beyond yolo v1. The improvement wherein the main points:
(1) Multi-scale increase the robustness of feature map, detecting both large objects, but also the detection of small objects, the accuracy increases;
(2) using whole Yolo connected to predict, but the SSD convolutional ensure speed. .
Network architecture
1, base network for the VGG-16, feature extraction picture, follow-up and multiple convolution operation, and extracted at different scales feature map are also carried out convolution to obtain the predicted borders of each feature map, and will put them together, be together NMS.
2, for each feature map to get convolution been predicted frame, which is different from faster-RCNN sliding window, but also from the regression prediction yolo, really good
3 different scales to predict when the feature map, different anchor box size
calculation method 4, the training process is similar to faster R_CNN, the prediction target block RCNN consistent loss function is almost the same
to sum up
The whole idea is similar to SSD YOLO, want to end the problem, but on the training design including loss of function, similar to faster R_CNN, while another two are not, that is, multi-scale feature map, this way, ensure accuracy, speed and guarantee. .