Overview of TensorFlow DeepLab for semantic segmentation

Semantic segmentation is the task of understanding and classifying the content of an image at the pixel level. Unlike object detection, where a rectangular bounding box is drawn over multiple object classes (similar to what we learned about YOLOV3), semantic segmentation learns the whole image and assigns a class of the enclosed object to the corresponding pixels in the image. Thus, semantic segmentation can be more powerful than object detection. The foundational architecture of semantic segmentation is based on the encoder-decoder network, where the encoder creates a high-dimensional feature vector and aggregates it at different levels, while the decoder creates a semantic segmentation mask at a different level of the neural network. Whereas the encoder uses a traditional CNN, the decoder uses unpooling, deconvolution...