search

lyndonzheng / Synthetic2Realistic

[ECCV 2018]: T2Net: Synthetic-to-Realistic Translation for Depth Estimation Tasks
180 stars 42 forks source link

Resize images and labels while test on Kitti? #16

Closed lsongx closed 5 years ago

lsongx commented 5 years ago

Hello lyndonzheng, thanks for your code!

While testing (using test_model.py), the images and labels are resized according to the setting in base_options.py, via

img_target = img_target.resize([self.opt.loadSize[0], self.opt.loadSize[1]], Image.BICUBIC)
lab_target = lab_target.resize([self.opt.loadSize[0], self.opt.loadSize[1]], Image.BICUBIC)

I have several questions about this:

  1. The depth labels in Kitti dataset are sparse. Is it okay to resize via Image.BICUBIC?
  2. Are the results reported on the resized images and labels? Is this setting the same with other comparison methods in your paper?

Thanks!

lyndonzheng commented 5 years ago

@LcDog 1. The test_model.py just uses the image to predict the depth value, the resized depth value will not be used during the test. The evaluation code can be found in util/evaluation.py. We will resize the predicted depth to the orginal KITTI ground truth resolution. That means the groundtruth will not be resized.

  1. The setting for all comparison methods is same in our paper.

Besides, the kitti depth is sparse, but during the training, we will first use the traditional methods to get the completed depth. The completion code can be found in monodepth (https://github.com/mrharicot/monodepth/blob/master/utils/evaluate_kitti.py)

lsongx commented 5 years ago

@lyndonzheng Thanks for your reply!!!