search

MingtaoGuo / yolo_v1_v2_tensorflow

Simple implementation of yolo v1 and yolo v2 by TensorFlow
MIT License
15 stars 4 forks source link

The question of the Yolo loss and the label #1

Open AugustusHsu opened 4 years ago

AugustusHsu commented 4 years ago

https://github.com/MingtaoGuo/yolo_v1_v2_tensorflow/blob/1ca4ff59e6c36bf7d92993aa8a2adc814863f366/utils.py#L99

https://github.com/MingtaoGuo/yolo_v1_v2_tensorflow/blob/1ca4ff59e6c36bf7d92993aa8a2adc814863f366/ops.py#L12

Do your label only consider one box per cell? What if the label has two or more boxes?

I have seen many of Github's codes only considering one box. Is this common? Or do I misunderstand something?

I'm not a native speaker of English, forgive for my grammar errors.

AugustusHsu commented 4 years ago

https://github.com/MingtaoGuo/yolo_v1_v2_tensorflow/blob/1ca4ff59e6c36bf7d92993aa8a2adc814863f366/ops.py#L24 https://github.com/MingtaoGuo/yolo_v1_v2_tensorflow/blob/1ca4ff59e6c36bf7d92993aa8a2adc814863f366/utils.py#L91

Hi, I have another question here. You use c_x / cell_w - col to normalize the coordinate. But, why not just use the original coordinate (after resizing the image to the 448x448)?