Dear fellow contestants,
After some initial difficulties, caused by the missing header/image_id, the organizers updated the files and I've made my first submission scoring less the 0.2 AP using YOLOv8 Nano
(I wonder why they separated the score for maize and bean instead of the two crop and weed classes)
During the training/validation of the later two submissions I've included the Rose dataset that I was able to include after translating the original XML annotations with segmentation into a COCO json with detection, reaching now an AP of almost 0.5 with almost 3000 images used overall for training (plus the augmented images) and still using a Small model for YOLO.
Just by viewing my test's resulting annotations (using Voxel51) I see very few cases where they could have been improved and I wonder if a higher score would necessary translate to sensible better performance on the field.
Do you think Object Detection holds real practical advantage over Segmentation when the time comes to do the real weeding (beside speed)?
What are you thought on this topic?
What models are you training on?
Are you willing to team up with me or include me in your team?
I hope this challenge is an opportunity for everybody to learn from each other and share interesting thought and insight on this subject, and if you think the same, please share your experiences here and/or let's team up!
(to the organizers: The team functionality is now enabled but still not working if you click on the Team tab)
regards
Eli