This leads to networks being designed to combine the information from earlier layers and high-resolution (low-level spatial information) with deeper layers and low-resolution (high-level semantic information).As we can see below, we first run our image through a standard classification network..We then extract features from each stage of the network, thus using information from a range of low-to-high..Each information level is processed independently before combining them all together in turn..As the information is combined, we upsample the feature maps to eventually get to the full image resolution.To learn more details about how segmentation with deep learning works, check out this article.The GCN Segmentation architecturePose EstimationPose estimation models need to accomplish 2 tasks: (1) detect keypoints in an image for each body part (2) find out how to properly connect those keypoints..This is done in three stages:(1) Extract features from the image using a standard classification network(2) Given those features, train a sub-network to predict a set of 2D heatmaps..Each heatmap is associated with a particular keypoint and contains confidence values for each image pixel about whether a keypoint likely exists there or not(3) Again given the features from the classification network, we train a sub-network to predict a set of 2D vector fields, where each vector field encodes the degree of association between the keypoints..Keypoints with high association are then said to be connected.Training the model in this way with the sub-networks will jointly optimise detecting the keypoints and connecting them together.The OpenPose Pose Estimation architectureEnhancement and RestorationEnhancement and restoration networks are their own unique beast..We don’t do any downsampling with these since what we are really concerned about is high-pixel / spatial accuracy..Downsampling would really kill this information since it would reduce how many pixels we have for spatial accuracy..Instead, all processing is done at the full image resolution.We begin by passing the image we want to enhance / restore to our network without any modification, at full-resolution..The network simply consists of a stack of many convolutions and activation function..These blocks are typically inspired and occasionally direct copies of those originally developed for image classification such as Residual Blocks, Dense Blocks, Squeeze Excitation Blocks, etc.. More details