search

JanTomesek / CrossLocate

CrossLocate
4 stars 0 forks source link

how your Alps semantic segmentation data is generated #1

Open ZQQ2019 opened 9 months ago

ZQQ2019 commented 9 months ago

Hello, reading through your code and paper, I have a couple of questions for you 1, how your Alps semantic segmentation data is generated by DEM model and OpenStreetMap, I'm interested in the details of this process! 2, in your original paper, you categorize the data set into train, val, test; but there is no test result for the test data set. 3，Your training model framework is tensorflow, have you considered using pytorch framework to train the model? I hope you can solve my problem as soon as possible in your busy schedule, I appreciate it!

JanTomesek commented 9 months ago

Hello,

  1. It depends on what exact details you are looking for, but have you taken a look at our Supplementary Material (http://cphoto.fit.vutbr.cz/crosslocate/data/supplement/CrossLocate-supp.pdf) ? In section 2.3, there is a brief description of the process. Simply said, it a 3D terrain model that is overlaid by 2D texture of landscape features and then rendered.
  2. That is not true as all the results (figures) in the paper are measured on the test sets, as clearly described in each figure or in the accompanying text.
  3. The implementation is very general and only lightly dependent on a particular ML framework. We chose TensorFlow but switching to PyTorch should be very straighforward, since the dependency on TensorFlow is quite small. However, it should not provide any particular benefit.
ZQQ2019 commented 8 months ago

Thank you for replying to me, I see your answer, I still have the following questions, I hope you have the time to solve the puzzle 1、Because of the short description of the production process of synthetic images in section 2.3 of the supplementary material and my relatively little knowledge of this piece, I am more interested in the detailed process and data processing, especially "it a 3D terrain model that is overlaid by 2D texture of landscape features and then rendered."

  1. I have reproduced the results of the study through your paper, but the programming file for generating the meta structures for the dataset is missing from your website (I have written one by myself through some fiddling).