ABSTRACT

To improve human action recognition performance of visual data, we proposed Hybrid Sampling Ensemble Learning method. 1)define the curvature distance distribution to fit discrete images into continuous expandable surfaces. 2)obtain density-weighted curvature distance of the images through class imbalance adjustment. 3)use auto-regressive classification tree strategy by utilizing the hierarchical regress of the filtered images. The classification performance of the ensemble learning method is not as good as the deep learning framework. However, the parameters are interpretable, and the construction of the classification framework is simple. In short, the proposed ensemble model is robust and controllable.

摘要

为了提高视觉数据对人类动作的识别性能，我们提出了混合采样合奏学习方法。1）定义曲率距离分布，以使离散图像适合连续的可扩展表面。2）通过类不平衡调整获得图像的密度加权曲率距离。3）通过利用滤波图像的分层回归，使用自动回归分类树策略。集成学习方法的分类性能不如深度学习框架好。但是，参数是可解释的，并且分类框架的构造很简单。简而言之，所提出的集成模型是鲁棒的和可控的。