Has anyone successfully trained on COCO?

[sbkim052]

Hello, thank you for sharing this code.
I used the code with no changes and when I saw the visualization of the generated boxes are not really good.

Has anyone successfully trained on COCO?

[yuexihang]

Hello. Which coco annotations did you use?

• 2017 Train/Val annotations [241MB]
• 2017 Stuff Train/Val annotations [1.1GB]
• 2017 Panoptic Train/Val annotations [821MB]

[sbkim052]

Hello @yuexihang
I used stuff_annotations_trainval2017.zip [1.1GB]
Thanks for the reply:)

[yuexihang]

Hello @sbkim052
The annotation stuff_annotations_trainval2017.zip [1.1GB] has only 92 kinds of categoris.
But the paper said they use the dataset with 171 annotations.

You can try to use the stuff and coco annotations together.

[JingyeChen]

Hi, according to the fifth page of this paper, it seems that the authors used the Panoptic version of COCO

[JackW987]

Hi, according to the fifth page of this paper, it seems that the authors used the Panoptic version of COCO

Hello,I use the panoptic version of coco,I modified part of the code in dataset.py to adapt to the corresponding version，the result is not right.