search

dvlab-research / PanopticFCN

Fully Convolutional Networks for Panoptic Segmentation (CVPR2021 Oral)
Apache License 2.0
391 stars 53 forks source link

Could you share the model on Cityscape as well? #4

Closed ChaoFang-TRI closed 3 years ago

ChaoFang-TRI commented 3 years ago

Thank you!

yanwei-li commented 3 years ago

Thanks for this suggestion. For the reason that some essential tools for Cityscapes (like panoptic dataset maker, instance-level random crop tools, etc.) have not been officially supported in Detectron2, we could not provide the training and inference code like a sub-project in Detectron2. We plan to release the full modified version and well as all the pretrained model later. Of course, it is easy to implement PanopticFCN for Cityscapes or other datasets with the currently released code.

lucasjinreal commented 3 years ago

@yanwei-li I am also interested on how panopticFCN preformance on cityscapes. It seems it's more convenient for cityscapes to train panoptic since it have mask png labeling already, however, detectron2 do not have such dataloader out-of-box to use. That would be very nice if you can publish a cityscapes trained model!