search

mseg-dataset / mseg-semantic

An Official Repo of CVPR '20 "MSeg: A Composite Dataset for Multi-Domain Segmentation"
MIT License
464 stars 79 forks source link

Invitation of making PR on MMSegmentation. #33

Closed MengzhangLI closed 3 years ago

MengzhangLI commented 3 years ago

Hey, here!

I am the member of OpenMMLab. This dataset and its related code/method are very valuable for segmentation. We hope we could introduce more people know this method and use dataset.

Would you like to make a new pr on MMSegmentation with us? We could work together to support this model and dataset effectively!

Best,

johnwlambert commented 3 years ago

Hi @MengzhangLI, thanks for your interest in our work. I'd be happy to make our work available to as many as possible.

Could you provide an exact design specification of what you would need in this PR? For example, the API design for training, dataset loaders, inference, taxonomy support, CPU/GPU support, and model storage?

johnwlambert commented 3 years ago

Hi @MengzhangLI, haven't heard back from you, so closing this issue. If you are able to put a design doc together with the implementation specification, I'm happy to chat more. Feel free to re-open the issue if you'd like to discuss more.

MengzhangLI commented 3 years ago

Hi, @johnwlambert

Sorry for late reply because first week of October in mainland China is holiday.

Comparing with model itself, it is easier to integrate dataset into MMSegmentation framework.

From our repo there are many widely used datasets. If it is official supported, we usually add its python file into ./config/_base_/datasets and ./mmseg/datasets, here is related pull request of adding COCO-Stuff dataset into MMSegmentation by our community member: https://github.com/open-mmlab/mmsegmentation/pull/625.

Let us keep in touch and hope we could work together to make more people use your dataset and model!

Best,

MengzhangLI commented 3 years ago

For model, it is a little bit complicated. I would spend some time reading your answer https://github.com/mseg-dataset/mseg-api/issues/34 to know more things.

Could you provide an exact design specification of what you would need in this PR? For example, the API design for training, dataset loaders, inference, taxonomy support, CPU/GPU support, and model storage?

For example, this is my re-implementation pr of BiSeNetV2, to add this model, I only do three things: (1) add model in ./mmseg/models/, (2) add config files in ./configs/ which MMSegmentation would use this file to build model, (3) add unit test file to control my pr quality during coding.

I think we only need to add model and dataset, don't need to care about dataloader and someother things if we could provide images and their pixel-wise ground truth.

Best,