key point of 《You Only Look Once: Unified, Real-Time Object Detection》

A single neural network in one evalution

Introduction

(1)simple
We reframe object detection as a single regression problem, straight from image pixels to bounding box coordinates and class probabilities.
(2)global image
unlike sliding window and region proposal-based techniques, YOLO sees the intire image during training and test time so it implicitly encodes contextual information about classes as well as their appearance.
and Fast R-CNN mistakes background patch because it can’t see the larger context.
(3)YOLO learn generalizable representations(概括性特征) of objects
a wide margin（大边缘）
(4)YOLO lags behind(落后于) current detection sys in accuracy

2 Unified Detection

entire image ->bounding boxs
predicts all bounding box across all classes for an image simultaneously
end to end
define confidence as Pr($object$)*IOU
if Pr(object) = 0; confidence = 0
if Pr(object) = 1; confidence = IOU(intersection over uion)
IOU用来表示预测框与真实框重合的程度，其最小值是0，最大值是1。 具体算法请参照如下文章

http://blog.csdn.net/u011534057/article/details/54845298

each bounding box consists of 5 prediction:x,y,w,h,confidence
each grid cell also predicts C conditional class probabilities, Pr($Class_i|Object$), predict one set of probabilities per grid cell, regardless of the number of boxes B

At test time we multiply the conditional class probabilities and the individual box confidence prediction
Pr($Class_i|Object$)*Pr($object$)*IOU =
class-specific confidence scores for each box
Question: the map of the box confidence and the grid cell
the center of the box in the cell

Model:
(1)detection as regression problem
(2)image ->divide S*S cells
(3)cell ->B bounding boxes
(4)the predictions are S*S(B*5+C) tensor

parameters: for YOLO on PASCAL VOC S = 7, B =2. The data had 20 labels classes, so C = 20. the prediction is 7*7*(5*2+20)