6. SSD model architecture in Keras
Unlike the code examples in the previous chapters, the tf.keras implementation of SSD is more involved. In comparison to other tf.keras implementations of SSD, the code example presented in this chapter focuses on explaining the key concepts of multi-scale object detection. Some parts of the code implementation can be further optimized such as caching of ground truth anchor boxes classes, offsets, and masks. In our example, the ground truth values are computed by a thread every time an image is loaded from the filesystem.
Figure 11.6.1 shows an overview of code blocks that comprise the tf.keras implementation of SSD. An SSD object in ssd-11.6.1.py builds, trains, and evaluates an SSD model. It sits on top of SSD model creator with the help of model.py and resnet.py and a multi-threaded data generator in data_generator.py. SSD model implements the SSD architecture as shown in Figure 11.5.1. The implementation of each major block will be discussed...
