wvangansbeke / MaskDistill

Discovering Object Masks for Unsupervised Semantic Segmentation [2022]
https://arxiv.org/abs/2206.06363
Other
93 stars 5 forks source link

VOC2012 instance segmentation annotation file in a COCO format #1

Open noelshin opened 2 years ago

noelshin commented 2 years ago

Hi @wvangansbeke, thank you very much for the great work on unsupervised segmentation!

Based on my understanding, you evaluated MaskDistill on the VOC2012 trainval set in COCO-style metrics using the detectron2 library, which requires a ground-truth annotation file in COCO format.

If so, could I get the ground-truth instance segmentation file of VOC2012 in a COCO format that you might use for the experiment? (I noted that VOC2012 annotation files come in a xml format which is not directly usable for a COCO-style evaluation.)

Many thanks in advance, noel

wvangansbeke commented 2 years ago

Hi,

Sorry for the late reply. Yes, you're correct. This was a personal preference. I converted the PASCAL VOC masks to the polygon format and the bounding boxes to xywh format myself. You should be able to find existing code online to do the conversion. If not, let me know. I'm still waiting on the final permission to release everything (this includes the ground-truth files).

noelshin commented 2 years ago

Thanks a lot for your reply - it'd be really great if I could get your script to convert VOC masks to a COCO format to ensure a fair comparison, if this doesn't interfere with your schedule on releasing your code and files. (Please feel free to send it to my email (gyungin@robots.ox.ac.uk) if that's better.)