SparsePose Training with Differentiated Heatmap Supervision for Improved Pose Estimation

Therefore, it is possible to train our proposed SparsePose directly in an end-to-end manner by utilizing a pose heatmap loss, similar to the approach adopted in numerous prior studies [5, 35, 52, 58, 62]. However, considering our comprehensive analysis of acquiring multiple heatmaps for accurate pose estimation, it becomes essential to explore the potential benefits of introducing differentiated supervision at the heatmap level. Hence, investigating this aspect can yield valuable insights and advancements in the field.