本文为论文的阅读笔记,并且参考了https://wenku.baidu.com/view/1dd6f253640e52ea551810a6f524ccbff121cab2.html
融合不同层的卷积响应,并产生高阶特征
instead of part annotations and explicit appearance modeling, we straightforwardly exploit the higher-order statistics from the convolutional activations.
we provide a perspective (观点) of matching kernel to understand the widely adopted mapping and pooling schemes on convolutional activations in conjunction with (结合) linear classifier.
designing an appropriate CNN architechture that can be plugged with non-linear local kernels in an end-to-end manner
With the kernel scheme, we extend our framework for higher-order integration of hierarchical (层次) convolutional activations. The effectiveness of fusing hierarchical features in CNNs has been widely reported in visual recognition. The benefits come from both the different discriminative capacities of multiple convolutional layers and the coarse-to-fine object description. However, the existing methods simply concatenate or sum multiple activations into a holistic representation, or adopt a decision level fusion to combine side-outputs from different layers. These methods, however, are limited in exploiting the intrinsic (内在的) higher-order relationships of convolutional activations in either the intra-layer level or the inter-layer level.
By using the kernel fusion on hierarchical convolutional activations, we can construct a richer image representation for cross-layer integration.
参考资料
https://wenku.baidu.com/view/1dd6f253640e52ea551810a6f524ccbff121cab2.html