人体骨骼点检测：自顶向下

1. Datasets

2. Human Pose Estimation

Top Down:自顶向下 → 先找人 后找点
Bottom Up:自底向上 → 先找点 后归纳

3. Keypoints Evaluation Metric

Precision measures how accurate is your predictions. i.e. the percentage of your predictions are
correct.
Recall measures how good you find all the positives. For example, we can find 80% of the
possible positive cases in our top K predictions.

真实关节点的格式：[𝑥1, 𝑦1, 𝑣1, … 𝑥𝑘, 𝑦𝑘, 𝑣𝑘]
Coordinates [𝑥, 𝑦] visible: 𝑣 𝑣 = 0: 未标注点
𝑣 = 1: 已标注但不可见
𝑣 = 2: 已标注并图像可见
𝑑i 是标注和预测关节点之间的欧式距离

4. Standard Deviation per Keypoint Type

5. Human Pose Estimation: Top Down

人体目标检测:
One Stage: Yolo, SSD等;
Two Stages: Faster RCNN, Mask RCNN等;
Anchor Free: CornerNet, CenterNet;
关键点检测：

6. FPN：Feature Pyramid Network

C1-C5的特征图尺寸是不同的；
使用1×1卷积确保depth一样先得到P5，
然后上采样确保特征矩阵大小匹配，
再进行特征矩阵相加，如法炮制得到P4,P3,P2。