Ball detection & tracking with a simplified CenterNet network
Автор: Antoine Keller
Загружено: 2023-03-12
Просмотров: 458
Описание:
Detecting small objects like sport balls in an image can be quite challenging for some networks that are used to work with big objects. I designed and trained a network based on the CenterNet architecture without regressing bounding boxes and without unnecessary up-sampling layers since tiny objects have very local features. This makes the architecture quite simple and training/inference quite fast. Goal was here to provide an accurate and reliable ball localization. It runs faster than real-time on my NVIDIA RTX A2000.
00:00 Single object tracking (soccer)
00:55 Output of the model (center and offsets heatmaps)
01:22 Multiple object detection (snooker)
See my Github repo for more details https://github.com/antoinekeller/ball...
CenterNet official implementation
https://github.com/xingyizhou/CenterNet
Resources for this video were taken from:
DFL Kaggle challenge https://www.kaggle.com/competitions/d...
Snooker • Bird's eye view of Ronnie's super-quick ce...
Thanks for watching!
Повторяем попытку...
Доступные форматы для скачивания:
Скачать видео
-
Информация по загрузке:
