Deep Linear Discriminant Analysis on Fisher Networks: A Hybrid Architecture for Person Re-identification

Abstract: Person re-identification is to seek a correct match for a person of interest across views among a large number of imposters. It typically involves two procedures of non-linear feature extractions against dramatic appearance changes, and subsequent discriminative analysis in order to reduce intra- personal variations while enlarging inter-personal differences. In this paper, we introduce a hybrid architecture which combines Fisher vectors and deep neural networks to learn non-linear representations of person images to a space where data can be linearly separable. We reinforce a Linear Discriminant Analysis (LDA) on top of the deep neural network such that linearly separable latent representations can be learnt in an end-to-end fashion. By optimizing an objective function modified from LDA, the network is enforced to produce feature distributions which have a low variance within the same class and high variance between classes. The objective is essentially derived from the general LDA eigenvalue problem and allows to train the network with stochastic gradient descent and back-propagate LDA gradients to compute the gradients involved in Fisher vector encoding. For evaluation we test our approach on four benchmark data sets in person re-identification (VIPeR [1], CUHK03 [2], CUHK01 [3], and Market1501 [4]). Extensive experiments on these benchmarks show that our model can achieve state-of-the-art results.

• The paper presents a hybrid architecture that integrates LDA with Fisher networks for improved discriminative feature extraction in person re-identification.
• It employs a modified LDA objective to ensure low intra-personal variance and high inter-personal differences in deep feature representations.
• Experimental results on benchmarks like VIPeR, CUHK03, and Market-1501 demonstrate its state-of-the-art performance and computational efficiency.

Analysis of "Deep Linear Discriminant Analysis on Fisher Networks: A Hybrid Architecture for Person Re-identification"

In the field of computer vision, person re-identification is a critical task involving the identification of individuals across different camera views. This complex problem demands a robust, discriminative feature extraction process to cope with appearance changes and optimize the matching accuracy. The paper "Deep Linear Discriminant Analysis on Fisher Networks: A Hybrid Architecture for Person Re-identification" by Lin Wu, Chunhua Shen, and Anton van den Hengel, proposes a novel hybrid architecture that addresses these challenges by leveraging both Fisher vectors and deep neural networks.

The authors introduce an innovative framework that prioritizes the generation of linearly separable feature representations in the latent space by incorporating a Linear Discriminant Analysis (LDA) into a deep network design. This is accomplished through a hybrid structure combining Fisher vectors with supervised neural layers, aiming to effectively minimize intra-personal variations and maximize inter-personal differences.

Methodology

The paper outlines a methodological framework where Fisher vectors, known for their efficacy in encoding local visual features, are integrated within a neural network. This hybrid architecture has an automatic feature learning capacity due to its deep network component. The network deploys a modified LDA objective function, ensuring that the learned feature distributions exhibit low variance within classes and high variance between classes. The approach is designed to be trained end-to-end using stochastic gradient descent, allowing the back-propagation of LDA-derived gradients, enhancing Fisher vectors' feature encoding capabilities.

Experimental Validation

The authors rigorously validate their approach using major person re-identification benchmarks: VIPeR, CUHK03, CUHK01, and Market-1501 datasets. Results indicate that their model achieves state-of-the-art performance, and the architecture significantly improves classification outcomes by ensuring linear separability in deep latent space, as evidenced by superior results compared to traditional metric learning and recent deep learning based re-identification techniques.

Implications and Future Directions

The hybrid approach proposed in this paper demonstrates notable advantages in computational efficiency and accuracy in scenarios with limited data, a common issue in person re-identification tasks. This architecture not only exemplifies improved class separability but also presents a feasible alternative to convolutional neural networks, especially in environments constrained by training data availability.

In terms of future developments, extending this framework to accommodate streaming data or adapting it for on-device implementations presents an attractive opportunity for research. Further exploration could focus on dynamic architectural adjustments that can optimize real-time re-identification tasks without sacrificing the model's linearly discriminative performance. Additionally, studies might explore adaptive learning mechanisms that respond to evolving data distributions typical in surveillance settings.

In conclusion, the integration of LDA within the deep network framework of Fisher vectors offers a compelling advancement in person re-identification methodologies, capitalizing on both unsupervised and supervised learning paradigms to enhance data representation's discriminative power. This hybrid architecture not only contributes to the theoretical understanding of deep learning applications in computer vision but also serves as a practical solution to one of its most challenging tasks.