1. 基于Pictorial Structures,即该结构包含不同部件之间的关系(比如头旁边是脖子),也包含不同部件之间的相对位置的概率(不同行人摆手的位置和手肘分布不同)
2. 增加人体部位的分割程度。将人体分割成12,24,51。分成24个不同部位的时候,效率最高。
3. 训练linear SVMs用dual coordinate descent。
4. PCP,PCK,APK不同的对预测点的评价方法。
PCP:A candidate body part is labeled as correct if its segment endpoints lie within 50% of the length of the ground-truth annotated endpoints.
PCK:A candidate keypoint is correct if it falls within a*max(h;w) pixels of the ground-truth keypoint, where h and w are the height and width of the bounding box respectively, and a controls the relative threshold for considering correctness.
APK:Average precision of keypoints (APK).Calculate the average of all PCK for all keypoints.