COCO-Stuff Dataset Segmentation


(Austin) #1

Hi, I’m curious if anyone here has trained the Dynamic U-Net on the COCO-Stuff dataset? I’m thinking about attempting it, but before I spend all the time/money I figured I’d ask if anyone else has already done so and would be willing to share weights.

Also if anyone has already figured out how to use the DataBlock API to read in the stuff_train2017.json and stuff_val2017.json segmentation annotations with SegmentationItemList, an example for that would be really helpful for me. Currently trying to get to there from the starting point of the datablock object detection data example which uses a json file. Wondering if I need to create image masks for each image as an intermediate stage or if I can use the json RLE segmentation annotations directly?


#2

Hi Austin,
Good question, and I was thinking about the same thing.
For the object segmentation, but the process would be the same.
I’ll let you know what I find, or if somebody else will get there first…