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

No correct bounding boxes #300

Closed Cbanyungong closed 6 years ago

Cbanyungong commented 6 years ago

the loss is always declining, but the accuracy dose not rise. When detecting Blood Cell,confidences of all boxed are 0.5 there is no correct boxes.Could you tell me what my mistake is?And the loss of WH is very high.

Loss XY [0.0567559749] Loss WH [5.61224461] Loss Conf [0.133079857] Loss Class [0] Total Loss [5.80208063] Current Recall [0.937499881] Average Recall [0.892882764]

cell

knightlyw commented 6 years ago

me too, I think the loss function is wrong, it's different from the formula

Cbanyungong commented 6 years ago

@knightlyw Have you solved it?

knightlyw commented 6 years ago

@Cbanyungong Don't do average

Cbanyungong commented 6 years ago

@knightlyw Thanks for your advice,I got the correct results

jmaity commented 6 years ago

@knightlyw Can you please help how to make the loss function correct? I'm facing the same issue.

E-LEAF commented 5 years ago

@knightlyw I'm facing the same problem. Can you tell how to avoid averaging in the loss function

kashyap-dhriti commented 5 years ago

@Cbanyungong how did you correct the results?