# Key Point Extraction the best method?

I have been researching about determining some key points on an image, in this case I'm gonna use cloth (top side of human body) pictures. I want to detect some corner points on those.

Example:

I have two solutions on my mind. One CNN with transpose layers resulting in heatmap where I can get points. The second is to get 24 number as output from the model meaning 12(x,y) point. I don't know which one will be better.

In face point detection, they use the second method. In human pose estimation, they use method one. So what do you suggest me to use? or do you have any new ideas? Thanks

• Have you tried to understand why, in face detection, they use the 2nd method, but, in human pose estimation, they use the 1st? Have a look at the literature. I am not familiar with one or the other method, but, if you have time, you could try both.
– nbro
Feb 7, 2019 at 17:39
• I did some reasearch but I currently dont know why they use the different methods. Maybe in heatmap you can detect more than 1 person. I will keep reaserching about this topic and specifically the difference between both methods. Feb 8, 2019 at 11:46

The 2nd method would make sense only if your object is at the same position in all test images. You would have such situation if you operated on crops located by a separate object detection algorithm. This happens to be the case in facial key-point detection.

The 1st method would be much more robust to various object poses since it is translation covariant by design. If a keypoint is detected at location A, it will be equally well detected at any other position with the same set of parameters.

• So, you mean that there are dense layers in the second method which cant work the same way for all points (to do that we need a lot data). Because heatmap method will be consist of only Conv layers, it will be better at generalization performing well in many unseen cases which are at different locations. Did I understand you well? Feb 8, 2019 at 19:58
• Yes. If the data are unaligned, the second method would have to learn more than the first method. It would therefore need more training examples to achieve the same generalization accuracy. Feb 9, 2019 at 14:31