search

imsoo / fight_detection

Real time Fight Detection Based on 2D Pose Estimation and RNN Action Recognition
MIT License
193 stars 42 forks source link

values and units of features #9

Closed stan-lu closed 4 years ago

stan-lu commented 4 years ago

What's the unit of Δpoint? Is it presented as number of pixels? If so, may I ask why it is not normalized by the width and height of the video? I believe it will be more robust if it is irrelevant to the size of the video. And I assume that the units of angle and Δangle are degrees but not rads. Is that correct?

imsoo commented 4 years ago

@stan-lu

Thank you for your interest in this repo.

What's the unit of Δpoint? Is it presented as number of pixels?

It is the distance in 2D position. (From the previous postion to the current position)

If so, may I ask why it is not normalized by the width and height of the video? I believe it will be more robust if it is irrelevant to the size of the video.

It just didn't occur to me. I agree with your idea. I think it's also a good idea to normalize with each person's height.

And I assume that the units of angle and Δangle are degrees but not rads. Is that correct?

Yes, it is degree.

stan-lu commented 4 years ago

@imsoo Got it. Thank you! I will try changing my video to the size of that in your training dataset to see if the result will be better. Is that 640 x 480?

imsoo commented 4 years ago

@stan-lu

I'm not sure what the result will be.

The size of the videos for training data set is as follows.

MHAD Dataset : 640 x 480 CMU Panoptic Dataset : 720 x 480