# Fisher Kernel for Deep Neural Activations

arXiv: https://arxiv.org/abs/1412.1628 · HTML: https://mlsanity.com/paper/1412.1628/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 4 Dec 2014.

## Built on by (most cited first, 1 of 1)

- [ASK: Adaptively Selecting Key Local Features for RGB-D Scene Recognition](https://mlsanity.com/paper/2110.07703/) (2021)

## Abstract (arXiv)

Compared to image representation based on low-level local descriptors, deep neural activations of Convolutional Neural Networks (CNNs) are richer in mid-level representation, but poorer in geometric invariance properties. In this paper, we present a straightforward framework for better image representation by combining the two approaches. To take advantages of both representations, we propose an efficient method to extract a fair amount of multi-scale dense local activations from a pre-trained CNN. We then aggregate the activations by Fisher kernel framework, which has been modified with a simple scale-wise normalization essential to make it suitable for CNN activations. Replacing the direct use of a single activation vector with our representation demonstrates significant performance improvements: +17.76 (Acc.) on MIT Indoor 67 and +7.18 (mAP) on PASCAL VOC 2007. The results suggest that our proposal can be used as a primary image representation for better performances in visual recognition tasks.