search

bowenc0221 / panoptic-deeplab

This is Pytorch re-implementation of our CVPR 2020 paper "Panoptic-DeepLab: A Simple, Strong, and Fast Baseline for Bottom-Up Panoptic Segmentation" (https://arxiv.org/abs/1911.10194)
Apache License 2.0
585 stars 117 forks source link

about input possible bug? #36

Closed pingjun18-li closed 4 years ago

pingjun18-li commented 4 years ago

Hi, bowen, In segmentation\data\datasets\base_dataset.py, line 119-121, dataset_dict['semantic'] directly uses label, where the values are original id (not trainId). This seems not consistent with the dataset_dict['semantic'] value set in line 127 (in fact assigned inside self.target_transform), where trainId is used. Is this a bug? Of course this has no effect for the panoptic dataset because hasInstance is set to True, e.g. in class CityscapesPanoptic. Also, I wonder what is the case that line 191-121 will be used.

bowenc0221 commented 4 years ago

There is a Cityscapes semantic segmentation dataset that will load semantic trainid labels, you can use this dataset to train semantic segmentation model only (e.g. DeepLabv3 and DeepLabv3+).

Originally I want to make it a segmentation toolbox that support both semantic and panoptic segmentation, but I'm just too lazy...