Accelerate box labeling

[ShinjoWang]

• Change for-loop iou calculation to once
• Change for-loop get-offset to once
• Change number of classes in nms

In my case (Intel(R) Core(TM) i7-6850K CPU @ 3.60GHz, NVIDIA GeForce GTX 1080Ti) -> Time per step reduced from 15.3s/step to 0.73s/step

[EParisot]

This is extremely usefull thankyou !