Introduction
This document serves as a formal correction to the paper "Action recognition with spatio-temporal augmented descriptor and fusion method," originally published in Multimedia Tools and Applications (2016). The correction addresses a mathematical inconsistency in the fusion kernel definition, which is a critical component for the integration of spatio-temporal descriptors in human action recognition tasks.
Corrected Mathematical Framework
The original publication contained a typographical or conceptual error in Equation 7, which defines the mechanism for fusing various feature sets. In action recognition, different descriptors capture distinct spatial (appearance) and temporal (motion) characteristics. To integrate these effectively, a weighted fusion approach is employed.
The corrected Equation 7 defines the fusion kernel $K$ for two video segments $x_i$ and $x_j$ as follows:
$$K(x_i, x_j) = \sum_{k=1}^{m} d_k K_k(R_i^k, R_j^k)$$
In this formulation:
- $R^k$ represents the $k$-th sub-descriptor.
- $d_k$ is the weight assigned to the $k$-th descriptor, reflecting its relative importance or informativeness.
- $K_k$ is the sub-kernel corresponding to the $k$-th feature set.
To maintain mathematical consistency between feature vector scaling and kernel fusion, each sub-descriptor $R^k$ is effectively scaled by the square root of its weight ($\sqrt{d_k}$). When the dot product is calculated, this results in the linear weighting by $d_k$ shown in the final kernel equation. Furthermore, a normalization constraint is applied to the weights to ensure the fused kernel remains properly scaled:
$$\sum_{k=1}^{m} d_k = 1$$
Significance and Impact
This correction is vital for the reproducibility and mathematical integrity of the research. The fusion method described is a form of Multiple Kernel Learning (MKL), a standard technique in machine learning for integrating diverse feature sets. By ensuring the fusion mechanism adheres to the requirements of a valid Mercer kernel, the stability and theoretical soundness of kernel-based classifiers (such as Support Vector Machines) used in the original study are preserved.
While the erratum does not alter the experimental results or the datasets used in the original work, it provides the necessary mathematical foundation for researchers attempting to implement or extend the spatio-temporal augmented descriptor framework. This ensures that the relationship between feature scaling and kernel weighting is correctly represented for future implementations.
Summary updated