Train on cityscapes

[lucasjinreal]

How to train on cityscapes dataset? Does any one trained on it?

[fmassa]

Training on a new dataset can be done by following the instructions in #15 , and by using the coco-like dataset format for cityscapes, which can be found in https://github.com/facebookresearch/Detectron/blob/master/tools/convert_cityscapes_to_coco.py
Alternatively, you can write your own dataset for that.

Also, in #166 @ltnghia managed to train on cityscapes.
Maybe he can share the procedure with us, and maybe (even better) send a PR adding support for CityScapes?

[lucasjinreal]

@fmassa Let me mention him there

[ltnghia]

@fmassa: I cannot click on "Create pull request".
@jinfagang: You can use my code in #166, it's correct. Convert Cityscapes to COCO format and then follow the procedure in #166, you definitely can train on Cityscapes.

[fmassa]

@ltnghia this is weird. Did you push the code to any public branch of your fork of mask RCNN benchmark?

[fmassa]

Fixed via #232