Table of contents
1. Classic algorithm and stage (Stage)
2. Ground Truth, Prediction and IoU
3. Precision and Recall
4. Confidence and Confidence Threshold
Five, AP (Average Precision), mAP (Mean Average Precision), all point interpolation method (Interpolation Performed in all Points) and AUC (Area Under Curve)
6. YOLO V1
Seven, non-maximum suppression (Non-maximum Suppression)
8. YOLO V2
Reference Paper
1. Classic algorithm and stage (Stage)
Two-stage (two stages): Faster R-CNN Mask R-CNN
Features: relatively slow speed, good effect, not suitable for real-time detection
One-stage (single stage): YOLO
Features: fast speed, relatively poor effect, suitable for real-time detection
2. Ground Truth, Prediction and IoU
3. Precision and Recall
Example: In the dimensioning of general technology, it is necessary to ensure both accuracy (no wrong marking, multiple markings) and recall rate (no missing markings)
4. Confidence and Confidence Threshold
Confidence > confidence threshold, it is marked as a positive sample; otherwise, it is marked as a negative sample
The higher the confidence threshold, the lower the recall and the higher the precision; the lower the confidence threshold, the higher the recall and the lower the precision
Five, AP (Average Precision), mAP (Mean Average Precision), all point interpolation method (Interpolation Performed in all Points) and AUC (Area Under Curve)
Adjust the confidence threshold and draw the PR diagram (including the red dotted line) of all point interpolation methods, as shown in the left figure below
AP calculation method: AUC,
As shown in the upper right figure,
6. YOLO V1
Network Architecture
Input image 448x448x3: C3, H448, W448
Output data 7x7x30 (5x2 (2 is the number of bounding boxes contained in each grid cell) + 20): GridX (1~7), GridY (1~7), (x1 (0~1), y1 (0~1 ), w1, h1, Confidence1, x2 (0~1), y2 (0~1), w2, h2, Confidence2, p(1), p(2),...,p(n=20)(n number of classes for the dataset))
输出数据量计算公式（SxS：网格单元数量，B：每个网格单元含有的边界框数量，5：归一化后的每个网格单元内的边界框横、纵坐标+边界框宽度、高度+边界框置信度，C：数据集分类数量）：
损失函数
位置误差
置信度误差
Lamda(noobj)=0.5：由于背景在图像中占比较大，故减小负样本（背景）影响，使正样本易于检测
分类误差
MSE（均方差）与CEE（交叉熵）：MSE适用于线性回归预测数值，即回归问题模型；CEE适用于逻辑回归测概率，即分类问题模型
此处应选择交叉熵作为损失函数
评价
优点：简单快速
缺点：每个grid cell仅预测一个类，多类重叠问题无法解决；小目标检测效果一般；当同一类出现的不常见的长宽比和其他情况时泛化能力偏弱
七、非极大值抑制（Non-maximum Suppression）
八、YOLO V2
提升概览
加入批归一化
提高分辨率
输入大小
V2为什么可以使用不同大小的输入：去掉了全连接层，解放了大小限制
网络结构
输入大小
为什么实际输入为416*416而不是448*448：416=13*32，我们期望输入大小（416）为13的奇数倍，使数据具有实际中心点
为什么网格单元数量由V1的7*7变为13*13：每个网格单元box数量有限，增加了网格单元数量从而增加了box数量
Darkent类型：Darknet-19，含有19个卷积层
3x3卷积：参考了VGG，参数较少，感受野较大
1x1卷积：效果与3x3相近，但节省了大量参数（偷工减料）
聚类提取先验框
先验框比例
如何确定先验框比例：利用数据集的标注信息，采用K-means划分先验框比例，k=5即分为五类
为什么k=5：Avg IOU较大，Clusters较小
距离的定义
为什么不用欧氏距离：防止误差随先验框大小变化而产生明显变化
Reference Paper：You Only Look Once: Unified, Real-Time Object Detection
