search

vcg-uvic / sportsfield_release

Code release for WACV 2020, "Optimizing Through Learned Errors for Accurate Sports Field Registration"
Other
69 stars 25 forks source link

Data augmentation #1

Closed jiangwei221 closed 3 years ago

jiangwei221 commented 4 years ago

Edit: We've update the arXiv version. Please refer to the newer arXiv version for more details.

Due to the page limit, we didn't include the data augmentation we used in our method. Besides the random perturbation, we mainly use 3 data augmentation methods.

  1. Random crop. We randomly crop 90% to 100% of the original image size.
  2. Horizontal Flip.
  3. Add random shadow. By observing the dataset, we noticed that many images contain the shadow of the stadium. This is done by randomly overlaying a black patch on to the image.

Method 1 and 2 require the correct caculation of the corresponding homography.

Hope this post can help you!