Converting ins_seg to coco, formatting issues about bdd100k

[pcicales]

Hello @XiaLiPKU sorry for the late reply, I wanted to make sure I wasnt making a mistake. The issue has nothing to do with how your json files are formatted (although there were some key differences that I had to change in order to match their requirements exactly, without deleting the extra entries). If anyone is interested, here is the thread response where I outline my solution, and what is likely the cause of the problem, which is rooted in a pytorch bug that has persisted for several releases. Thank you for your responsiveness and help!

[XiaLiPKU]

Sorry, the doc here really lack something.
The correct command is here:
python3 -m bdd100k.label.to_coco -m ins_seg|seg_track --only-mask -i ${mask_base} -o ${out_path}
As you only feed bitmasks as the input, the --only-mask is required.

[XiaLiPKU]

The issue I am having with this annotation set is that it appears to be causing issues in my other code (which I believe is due to the scalabel content)

Can you paste the error report here?

[XiaLiPKU]

More to say. If you want to generate coco with mask in polygon format instead of RLE, you can add --mask-mode polygon to the command line

[pcicales]

Here is a link to the thread I am posting on regarding this isssue.

I am also running your solution now to see if it solves things.

[pcicales]

Hi @XiaLiPKU , going through the json files and comparing to the coco formatting guidelines, I am having trouble understanding a few of the elements saved after running the to_coco script correctly. For example, it seems that type and video are not compatible with coco. Perhaps I am misunderstanding something, as I thought that these elements were retrieved using these keys (making it to where extra keys are not a problem). But I continue to have issues loading the data with pytorch assuming the data is formatted following the coco guidelines. Have you encountered this before?

[XiaLiPKU]

Indeed, we add some extra elements, and they are used for tracking problems (box_track & seg_track).
However, given my experience, these fields will not influence your usage of MMDet.

For the error reports, can you paste them here? Or you can upload the converted json files.

[guifentian]

More to say. If you want to generate coco with mask in polygon format instead of RLE, you can add --mask-mode polygon to the command line

@XiaLiPKU
I am trying to convert to polygon format using --mask-mode polygon as you suggested. but it was not recognized.

error: unrecognized arguments: --mask-mode polygon

how to do that?