Compression Techniques for Deep Fisher Vectors

Diversified Fisher kernel: encoding discrimination in Fisher features to compete deep neural models for visual classification task

“…Fisher kernels derived from DBMs [14] to achieve large‐scale visual classification task. We have observed that the standard practice of utilising Fisher kernels derived from deep neural models [14, 16, 19] has the following limitations: (i) results in large dimensional FVs that require more memory usage and disk space for classifier training, and (ii) prone to overfitting under limited data availability applications due to the large number of gradient parameters derived from deep neural models. Besides these shortcomings in deep FVs, there is also a growing need for deploying computationally efficient and compact features in resource constrained environments or applications such as mobile devices, smart wearables, autonomous vehicles and so on.…”

“…FVs for image classification have been improved in several different ways by introducing class‐relevant information through foreground Fisher kernels [20] and by utilising diagonal Fisher information matrix to approximate Fisher kernels [21, 22]. To address the problem of high dimensionality in FVs and find compact representations, several compression and feature selection methods have been evaluated before [16, 23]. The techniques, selection as well as compression, seek to reduce the dimensions of the data descriptors; however compression methods implement the technique by projecting the data in a different parameter space, whereas the feature selection methods carefully choose the data attributes without changing them and hence retain the original data space.…”

“…Peng et al also explored a similar idea for action recognition task. Ahmed and Azim [16] suggested to derive a different class of FVs from deep Boltzmann machine (DBM) and showed improvement in accuracy yielded by the Fisher kernel derived from restricted Boltzmann machine (RBM) on benchmark image classification data sets [17]. The classification accuracy obtained by the proposed framework was shown to be competitive to that of state‐of‐the‐art deep learning models such as deep belief nets and classRBM.…”

“…Chandrasekhar et al [25] studied the problem of global descriptor compression and showed how sparse high‐dimensional CNN representations can be hashed to very compact representations (64–1024 bits) for large‐scale image retrieval with little loss in matching performance. In order to reduce the space requirements of Fisher features for quick retrieval, FVs have also been binarised [16, 28, 29]; however this has shown to reduce the classification accuracy of the developed artificial vision systems.…”