YOLO-v1
- Prior work: classifiers
various locations and scales
** DPM** (deformable parts models) sliding window approach HOG +SVM
**R-CNN ** bounding boxes classifier–>post-processing to refine itslow and hard to optimize
Fast R-CNN
Fast YOLO: 155 frames per second - YOLO
single regression problem
separated bounding boxes and class probabilities
single neural network
be optimized end-to-end directly
45 frames per second
twice the mean average precision
more localization errors ,less to predict false positives on background
ACCURACY LAG
3.置信度,每个格子预测B个box,每个box有一个检测到目标的置信度,每个格子有一个分为各个类的置信度,IOU是分别B个box/实际box分别与8个box的并。
architecture
- 结构,24层卷积层和2层全连接层