search

experiencor / keras-yolo2

Easy training on custom dataset. Various backends (MobileNet and SqueezeNet) supported. A YOLO demo to detect raccoon run entirely in brower is accessible at https://git.io/vF7vI (not on Windows).
MIT License
1.73k stars 784 forks source link

Video issues #392

Open Bergo1305 opened 5 years ago

Bergo1305 commented 5 years ago

When I predict on live video (live football match) i got a big bounding box who covers my desktop with probabilty near 40%. So what I need to change to fix this issue? When camera is near player it works fine, when it is far away it behaves weird? Thank you.

MBoaretto25 commented 5 years ago

Well, does you dataset has images of the whole field while the camera is not near to the player?

It might be because your object size will change depending on how far is the player from the camera. You might wanna train for player detection with images when the camera is both far and near the player.