SegFix cannot be used in my own dataset? about openseg.pytorch HOT 16 CLOSED

Hi,
I'm really interested in using SegFix for my own dataset. The paper seems to have been released on arXiv, but do you have any plan to release the tutorial and the code to train SegFix for our own dataset? Or, if you've already released them, please let me know.

from openseg.pytorch.

Same interested in training SegFix on the CUSTOM dataset, could you please share some advice?

from openseg.pytorch.

Same question here. How to generate *.mat offset files for custom datasets?

from openseg.pytorch.

I have the same question, How to generate the offset files? the origin offset files are for cityscapes datasets. And the Segfix paper I can not find in google or arxiv or anywhere.

from openseg.pytorch.

Thanks for your interest in our work. It requires to train the SegFix model on your own dataset that is used to predict the offset files. Currently, we have not released our paper on the arXiv and we will release the SegFix paper depending on the ECCV result.

from openseg.pytorch.

@LPengC @kotetsu-n We would like to release the training code of the SegFix on the Cityscapes & ADE20K benchmark. Especially, our SegFix can be trained with a mixture of multiple different datasets. Please stay patient.

from openseg.pytorch.

@LPengC @kotetsu-n We would like to release the training code of the SegFix on the Cityscapes & ADE20K benchmark. Especially, our SegFix can be trained with a mixture of multiple different datasets. Please stay patient.

Another question, why use Sobel filter on the distance map can represent directions between boundary pixels and interior pixels? Any responses would be appreciated!

from openseg.pytorch.

Same here - looking forward for the release of segfix training code.

from openseg.pytorch.

@PkuRainBow Congratulations on the SegFix ECCV publication! I am looking forward to the release of the model and training code. Are you planning to release your pre-trained model weights also?

from openseg.pytorch.

@hsfzxjy will help to organize and release the training code and checkpoints of our SegFix method.

from openseg.pytorch.

@PkuRainBow @hsfzxjy Great to hear that you are planning to release checkpoints also.

@LPengC The Sobel filter computes (an approximation to) the derivative of the distance map. And the derivative of the distance to the boundary evaluated at an image coordinate will be a 2d vector pointing to the closest point on the boundary. This is used to generate the per pixel directions in the training data. At least that is my understanding.

from openseg.pytorch.

Thanks to @pocketpixels!
@LPengC Not exactly, the distance map records the distances between the pixel and the nearest boundary. We apply the Sobel filter to estimate an offset pointing to the position with the largest distance value given a candidate sliding window.

We will release the code on how to generate the ground-truth direction/boundary maps on Cityscapes soon.

from openseg.pytorch.

@pocketpixels @shgidi
We have just released the implementation of SegFix. Please refer to https://github.com/openseg-group/openseg.pytorch/blob/master/MODEL_ZOO.md#segfix-1 for more details.

For some reasons our result with Higher-HRNet is currently unavailable, and we provide checkpoints and script examples based on HRNet-W48. We will release them as soon as they are ready.

Thanks for your patience!

from openseg.pytorch.

@hsfzxjy Great, thank you!
Are you planning to still release the training code and checkpoints for ADE20k as well?

And do you think it could make sense to use a HRNet-W18 but with the Higher-HRNet trained transposed convolution at the end? Or would the W18 HRNet likely not have sufficient bandwidth for the higher output resolution?
For my application I want/need the higher output resolution, but am targeting mobile, so I am constrained in terms of memory use and computational complexity.

from openseg.pytorch.

@pocketpixels Currently, we have no plan to release the checkpoints for the ADE20K.

We have released the data loader to train the unified SegFix model on multiple datasets simultaneously. Please make the related training script public. @hsfzxjy

It might be reasonable to extend the HRNet-W18 following the https://github.com/HRNet/HigherHRNet-Human-Pose-Estimation while we have not verified such extension in the current stage.

We guess you might first verify the performance of the original HRNet-W18 and then try other advanced extensions.
If you want to further decrease the computation cost, you can also try the HRNet-W18-small.

Any new results on your own tasks are welcomed to be shared.

from openseg.pytorch.

@PkuRainBow Thank you very much for your reply.

from openseg.pytorch.