swamiviv / lsd-seg Goto Github PK

First thank you for sharing the great work. From previous issues you said we needed to map the labels of GTA V to CityScapes, could you pls upload the scripts to do that?

If for some reason the script is unavailable now, I want to ask whether we can directly use the code in the website https://download.visinf.tu-darmstadt.de/data/from_games/ readImage.m.

In the paper: about SYNTHIA dataset, you use 16 common classes with cityscapes,

why these places indicate that class number is 19?

https://github.com/swamiviv/LSD-seg/blob/master/code/train.py#L62

https://github.com/swamiviv/LSD-seg/blob/master/code/torchfcn/datasets/segmentation_datasets.py#L261

Hi

Has anyone validated the results on the given 3 datasets and then tried to test with other datasets ?

I have tried the implementation with some datasets I have available but there is no improvement in the F networks performance on Target dataset after training for few epochs. The generator and discriminator learn to produce realistic images though. Does changing the adv_weight in F network loss improve the chances of domain adaptation in practice?

Thanks in advance.

Please attach comment on the code related to IoU.
The following code is sort of hard to understand:

BTW, I got bad evaluation results on cityscapes dataset after training on GTA5 dataset.
While the final mean accuracy for validation on GTA5 is 54.8536%, we got the following evaluation results:

These are quite different from those listed in your paper:

Hi, could you please tell which exact version of SYNTHIA dataset you used for the experiments.
Is it SYNTHIA-RAND-CVPR16 or SYNTHIA-RAND-CITYSCAPES?
Thank you.

Hi, I run your code on GTA2Cityscapes but failed to reproduce the evaluations claimed in your paper. During the training, I also noticed that the F's loss is not optimized at all, meaning that the domain shift is not address well? Here is what I got and it is worse than sourceonly. Any ideas? Did I do something wrong?
===>road: 83.28
===>sidewalk: 28.78
===>building: 61.49
===>wall: 6.52
===>fence: 0.19
===>pole: 8.44
===>light: 2.17
===>sign: 0.48
===>vegetation: 58.63
===>terrain: 14.29
===>sky: 55.88
===>person: 18.45
===>rider: 0.11
===>car: 53.92
===>truck: 2.9
===>bus: 1.52
===>train: 0.0
===>motocycle: 0.12
===>bicycle: 0.03
===> mIoU: 20.91

Hi @swamiviv , thanks for your code. Your results of source only vgg16-fcn8s with GTA to cityscapes is 29.6, which is much higher than Curriculum domain adaptation with 22.0. What is the difference ?

Hi,

when I set the image size to be 1024,512 then there appears the out of memory error. I was using TITAN X gpu. Did you run into the same problem? how did you train with image size 1024,5112?

using default setting, the mIoU is 28%.

Does anybody else get better result?

I think the training of GAN is very tricky.

I can not find the synthia_mapped_to_cityscapes(labels) in dataset. Who can help me?

This happens when executing nll_loss located in code/torchfcn/utils.py under the mode of sourceonly. The training data is from GTA5.

The error occurs because the snippet inside the cross_entropy2d() first use a mask to exclude elements whose values are less than 0 in target(that is, labels). In other words, mislabeled pixels are not involved when calculating cross entropy.

However, the corresponding prediction values for those removed pixels still exist in log_p, which leads to the array size conflict.