YOLO: detection in one look
On 8 June 2015 Joseph Redmon, Ali Farhadi and colleagues at the University of Washington and the Allen Institute for AI described YOLO: one network in one pass predicts boxes and classes for the whole image. It runs at 45 frames per second, with 58.8 mAP on PASCAL VOC 2007.
Why it matters
Object detection became possible in real time on one GPU, about 100 times faster than Fast R-CNN, at the price of moderate accuracy. It put speed beside accuracy as a detector's goal.
On VOC 2012, 54.5 mAP, which the paper itself calls below the state of the art; small objects are the weakest. Combined with the best Fast R-CNN, accuracy on VOC 2007 rises by 2.9 points to 74.7%. Speed was measured on an Nvidia Titan X. The record does not claim Fast YOLO, 155 frames per second or 63.4 mAP: none of them is in the first version.