5. Results
In our experimental evaluation of our neural network,both training and testing were done on a Linux PC with an Intel Xeon E5-1620 CPU, two NVIDIA Tesla K20 G-
PUs and 32GB memory. For 10000 panoramas containing traffic-signs, we separated them into a training set and a testing set (as explained in the released benchmark), with about 2:1 ratio to give the deep learning methods plenty of training samples. The other 90000 panoramas were included during testing.
We used the evaluation metrics used for the Microsoft COCO benchmark, and divided the traffic-signs into three categories according to their size: small objects(area < 32 2 pixels), medium objects(32 2 < area < 96 2 ) and large objects (area > 96 2 ). This evaluation scheme can tell the ability of a detector on different sizes of object.