princeton-vl / pytorch_stacked_hourglass

Pytorch implementation of the ECCV 2016 paper "Stacked Hourglass Networks for Human Pose Estimation"
BSD 3-Clause "New" or "Revised" License
465 stars 94 forks source link

quarter shift calculation question #15

Closed dmitrysarov closed 3 years ago

dmitrysarov commented 3 years ago

https://github.com/princeton-vl/pytorch_stacked_hourglass/blob/64907788e4c69616aac9ab044ae26fc010ed89c3/utils/group.py#L41 As i understood there is no actually second peak search, the shift is just performed in the direction of skew of distribution, right?

crockwell commented 3 years ago

can you clarify what you mean by "second peak search"? As I understand, I think you are correct: the shifts are just small sub-pixel changes based on location in image, which is motivated by skew in distribution (or similar idea).

dmitrysarov commented 3 years ago

I mean - to search real local isolated peak, by mean, for example, https://scikit-image.org/docs/dev/auto_examples/segmentation/plot_peak_local_max.html

crockwell commented 3 years ago

No, we don't use a function like that.