Giters

dahiyaaneesh / peclr

This is the pretraining code for PeCLR. An equivariant contrastive learning framework for 3D hand pose estimation. The paper is presented at ICCV 2021.

Home Page:https://ait.ethz.ch/projects/2021/PeCLR/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

augmentation order

hongsukchoi opened this issue · comments

commented

Hi,

According to the comments you wrote here, it seems like there has to be a specific order for different augmentations.

Is there any reason for this?
If I change the order, is there bad effects on the accuracy?

commented

Yes,

This was done to ensure that artifacts in the transformed images are minimized. For example, if one crops the image first and then rotates the image by 30 degree with no scaling the image has artifacts on new corners. However rotating first followed by cropping can minimize them, as chances are high the cropping will remove the artifacts around the corner.

To answer your second question: No we did not evaluate the effects of different ordering on accuracy but I encourage you to try that. We did qualitatively evaluate the input samples with the order mentioned in the code. This yielded good input samples with negligible artifacts that could have led to poorer performance.

I hope it helps.