What is DeepLab and what is the idea of ASPP?

DeepLab is a neural network architecture for semantic image segmentation that effectively highlights objects at different scales. The main idea is to use dilated convolutions to increase the receptive field without losing resolution.

ASPP (Atrous Spatial Pyramid Pooling) is a key component of DeepLab, which applies several parallel convolutions with different dilation rates. This allows capturing context at multiple scales simultaneously, improving the segmentation quality of objects of various sizes.

Example: in ASPP, several convolutions with dilation=1, 6, 12, 18 are combined to obtain a rich contextual representation.