I am currently writing my thesis about human pose estimation and wanted to use Google's inception network, modify it for my needs and use transfer learning to detect human key joints. I wanted to ask if that could be done in that way?
Assuming I am having n-keypoints, generating the n-feature maps, use transfer learning and cut off the final classification layers and replace it by a FCN which guesses the key joints. I am asking myself if this might be possible.
However, these feature maps should output heatmaps with the highest probability as well. Is this assumption valid?