search

sud0301 / semisup-semseg

136 stars 30 forks source link

Inconsistent with the results of Hung et al. paper #1

Closed zhangbin0917 closed 5 years ago

zhangbin0917 commented 5 years ago

I have noticed that the there are inconsistent with the results of Hung et al. [15] paper. Did you re-implement their code?

sud0301 commented 5 years ago

Hi, All the results are consistent with the results of Hung et al. [15] paper. There is a difference in results for Cityscapes dataset since Hung et al. uses COCO pre-training and our results are produced without using COCO pre-training. For PASCAL-VOC dataset, we show both results with and without COCO pre-training. All the numbers for Hung et al. [15] are produced using their official repository.

zhangbin0917 commented 5 years ago

Thank you for your reply. I have another question. Have you run only one experiment for each number in the table? If it is yes, the result is very uncertain and not convincing. Should you set different seeds in different labeled data and unlabeled data to perform multiple experiments to calculate the mean and standard deviation?

sud0301 commented 5 years ago

Each number is evaluated using three different splits and the mean value is reported.