Hello! Great work on putting together such a useful and large dataset.
I was wondering if there was any talk of potentially providing the vector labels for each image in a format which can be readily ingested by detectron2 and/or mmdet. I'm thinking about for example if one wanted to train an instance segmentation with Mask-RCNN or other more bespoke models, or otherwise wanted to work on a model which isn't already built into torchgeo. Detectron2's guide to custom datasets require labels to be COCO-like, which from my recent experience is still fairly cumbersome and baroque requiring a fair amount of data wrangling. If any of you have experience converting GeoParquet to a COCO-like format, directing towards resources would be great, otherwise I will probably need to convert GeoParquet to GeoJSON and then to COCO.
In sum, any suggestions for the least painful way to getting the ag field labels into a format which can be plugged into a generic computer-vision framework would be great. Thanks in advance!
(ETA: if this would be more suited to the ftw-baselines Issues page, please move it there instead. Thanks again!)
Hello! Great work on putting together such a useful and large dataset.
I was wondering if there was any talk of potentially providing the vector labels for each image in a format which can be readily ingested by detectron2 and/or mmdet. I'm thinking about for example if one wanted to train an instance segmentation with Mask-RCNN or other more bespoke models, or otherwise wanted to work on a model which isn't already built into torchgeo. Detectron2's guide to custom datasets require labels to be COCO-like, which from my recent experience is still fairly cumbersome and baroque requiring a fair amount of data wrangling. If any of you have experience converting GeoParquet to a COCO-like format, directing towards resources would be great, otherwise I will probably need to convert GeoParquet to GeoJSON and then to COCO.
In sum, any suggestions for the least painful way to getting the ag field labels into a format which can be plugged into a generic computer-vision framework would be great. Thanks in advance!
(ETA: if this would be more suited to the ftw-baselines Issues page, please move it there instead. Thanks again!)