nilboy / tensorflow-yolo

tensorflow implementation of 'YOLO : Real-Time Object Detection'(train and test)
774 stars 312 forks source link

The coord loss always fluctuates and doesn't converge with training going on. #22

Closed pureview closed 7 years ago

pureview commented 7 years ago

I'm using your code to locate human eyes. In the first 500 iterations, coord loss drop down from 8.x to 1.x and it doesn't drop after that. Please provide some advice. Here is the test demo after 230k iterations. It's obvious that CNN could detect pupil but cannot locate it accurately.

368 470 a 20161115161959679 00