Evaluation metrics systematization for 2D human poses analysis models

This paper describes the systematization of evaluation metrics for 2D human pose analysis models. Some most popular tasks solved using machine learning (ML) methods are detection, tracking and recognition actions various practical applications. There a lot different that allow evaluating model from one point or another. To evaluate specific task, certain set is used. However, as literature shows, vast number metric definitions, well use terms multiple representations same ideas, causes problems interpretation comparison ML models in detecting, tracking, recognizing actions. The purpose this work to analyze processing poses video order facilitate informed choice metrics. improve objectivity results empirical studies existing newly developed actions, into subgroups was proposed, depending on what task they evaluate. Four classes were introduced: classification metrics, key point’s object general Classification based quality matching values predicted bounding boxes with ground truths. Key detection oriented found joints body skeleton. Tracking each frame correctness determining its trajectory. General not specifically related any tasks. prototype application suggested systematization, which help data scientists formalizing problem being area developed. demonstrate application, Faster R-CNN, SSD YOLOv3 analyzed compared scope area. showed R-CNN have accurate responses, although disadvantage high False positive rate. implementation also True negative uninformative working boxes, because inability calculate negatives image data.