Yuwen Xiong

@tlittletime @Bigwode @IbsenChan @itongworld I haven't tested the code yet, but I believe that is because the authors wrongly set momentum=1 for all batch norm layers, which makes the BN...

This codebase needs a slightly modification to work under python 2.7, if you don't know how to do that please simply use python 3.6/3.7

ResNet101 results use deformconv for res3-5

please follow this https://github.com/uber-research/UPSNet/issues/36#issuecomment-491593609

Unfortunately this codebase doesn't support multiple images on one gpu. You can try to run multiple horovod workers on one gpu to mimic it, then one gpu will effectively take...

deformconv is only used in semantic head in our provided setting, if you want use it in backbone you could set config.network.backbone_with_dconv to 3, 4, or 5 (means it will...

Please make sure you pull the latest code and use_syncbn is set to false in the config file that you use.

We generate gt_masks here: https://github.com/uber-research/UPSNet/blob/2ced987576a98dc16c5a580839b657948644d44d/upsnet/dataset/cityscapes.py#L179 and here: https://github.com/uber-research/UPSNet/blob/2ced987576a98dc16c5a580839b657948644d44d/upsnet/dataset/coco.py#L160 . It will only contain 0 and 1. For your second question, a basic assumption is that there is no overlapping between...

Thanks for pointing it out, I confirm that these two issues exist. The first one will generate wrong panoptic gt when image shape cannot be divided by 32 (which does...

Thanks for your attention, it's on the plan but may not very soon. Currently I only ran on instance segmentation task and got 21.9 mask mAP. Contributions on ADE20K are...