about datasets/prepare_coco_semantic_annos_from_panoptic_annos.py

[fuzhao123232]

thanks for your job. I have a question: why COCO Instance Segmentation need panoptic_annos in the eval stage?

[fuzhao123232]

if i don't have panoptic annos ,can i skip this eval step? Now I am training my custom datasets

[junjiehe96]

hi, thanks for your interest in our work. Training COCO Instance Segmentation (i.e., cfg.DATASETS.TRAIN=("coco_2017_train",) only requires instance annotations. If you want to training your own datasets, you can refer to this to register your datasets in detectron2 format, and then use our fastinst dataset mapper to process them.

[fuzhao123232]

大哥，因为你们这里要求了要运行这个脚本生成语义标签的，但是我没有语义标签，所以运行这个就报错了。我的数据集已经注册了，并且可以train和demo推理； 但是zh这是因为我注释了所有train里面eval的步骤 如果不注释的话，就会报这个错： 然后我的segm是这样的

[fuzhao123232]

我的标签仔细检查了没有问题，并且如果有问题的话，我的infer应该会很异常，但我的推理也是正常的

[fuzhao123232]

![Uploading image.png…]() 因为我的标签里面的iscrowd是=0的，所以标签都是polygon，train和demo都正常，但eval好像要求我的格式是rle，就很奇怪

[junjiehe96]

1. python datasets/prepare_coco_semantic_annos_from_panoptic_annos.py 用来全景分割，否则不是必须的；
2. 如您所说，可以尝试将polygon转成rle；