search

charusharma1991 / SSL_PointClouds

34 stars 5 forks source link

About the accuracy #9

Closed priceee closed 1 year ago

priceee commented 3 years ago

Thanks for your excellent work. I noticed that the accuracy in the paper is written in a range (like 81.50±2.25%). I wonder when you get these accuracy, did you use random classes in every time you run the experiments, or you just fix the classes? And how many times you repeated ?

priceee commented 3 years ago

In the code you only used split 1 of the data, when you testing your model, are you testing on split 1~5 or just split 1?

charusharma1991 commented 3 years ago

Thanks for your interest!

  1. We use random classes in every run and we consider 10 such different splits to get the avg results mentioned in the paper.
  2. We make different train-test splits. so, for one training set, we consider one test set.
priceee commented 3 years ago

Thanks for your interest!

  1. We use random classes in every run and we consider 10 such different splits to get the avg results mentioned in the paper.
  2. We make different train-test splits. so, for one training set, we consider one test set.

Thank you for your answer! I have another question in the segmentation tasks, in shapenet dataset there are 16 classes, however when running with 10-way-10-shot FSL learning, you can only randomly pick 10 classes. So how do you get the mIoU for each of the 16 classes when you can only randomly pick 10 classes every time?