feat: mdv5a coco data prep
Hi Dan,
Thanks for the quick review, updated the following. Do let me know if this looks good :)
- coco_download.py: Downloads train+val+test
- coco_data_prep.py:
- Takes in the datasetYear as arg (e.g. train2017)
- Filters person and vehicles images and annotations
- Converts COCO to YOLOv5 format
Looks great! You marked this as a draft and indicated "WIP" in your comment, so it looks like you're just looking for a quick review now, not a merge. The download script and Python script both ran without any issues, I see no concerns.
One quick note: although all of this data would be in the train split from MegaDetector's perspective, there's no reason not to use the train+val+test data from COCO, rather than just the train data. It doesn't change the code at all, just a couple more lines in the download script.
For my two cents, I would probably do the download in Python also, just to keep things all in one place, but the shell script is fine too.
Thanks!