A sparse contrastive decline is even further proposed to teach our design with sparse supervision. Our new formulation generates a 3D heatmap indicating the plausibility of all locale/scale mixtures in one network ahead go, which is more than ten instances speedier than the previous sliding-window process. What's more, it supports interactive searc