[Paper Review] Adaptive Re-ranking of Deep Feature for Person Re-identification

Yong Liu, Lin Shang+1 more|arXiv (Cornell University)|Nov 21, 2018

Video Surveillance and Tracking Methods25 references4 citations

TL;DR

This paper proposes Adaptive Re-Ranking (ARR) and Deep Feature Fusion (DFF) to enhance person re-identification by leveraging contextual information and sub-feature diversity. ARR iteratively increases contextual relevance during re-ranking, while DFF fuses sub-features via a diffusion process; together, they achieve state-of-the-art performance, with ARR-DFF outperforming SPreID by 4.2% in mAP on Market and 7.5% on Duke.

ABSTRACT

Typical person re-identification (re-ID) methods train a deep CNN to extract deep features and combine them with a distance metric for the final evaluation. In this work, we focus on exploiting the full information encoded in the deep feature to boost the re-ID performance. First, we propose a Deep Feature Fusion (DFF) method to exploit the diverse information embedded in a deep feature. DFF treats each sub-feature as an information carrier and employs a diffusion process to exchange their information. Second, we propose an Adaptive Re-Ranking (ARR) method to exploit the contextual information encoded in the features of neighbors. ARR utilizes the contextual information to re-rank the retrieval results in an iterative manner. Particularly, it adds more contextual information after each iteration automatically to consider more matches. Third, we propose a strategy that combines DFF and ARR to enhance the performance. Extensive comparative evaluations demonstrate the superiority of the proposed methods on three large benchmarks.

Motivation & Objective

• To improve person re-identification by better utilizing deeply-learned features.
• To address the limitation of existing iterative re-ranking methods that do not adaptively increase contextual information over iterations.
• To exploit the diverse information embedded in different parts of deep features for improved metric learning.
• To develop a robust, parameter-insensitive strategy combining re-ranking and feature fusion for enhanced re-ID performance.

Proposed method

• Adaptive Re-Ranking (ARR) iteratively re-ranks matches by dynamically expanding the contextual neighborhood structure to include more relevant samples.
• ARR uses a diffusion-based process on a similarity graph to propagate affinity scores, increasing contextual information with each iteration.
• Deep Feature Fusion (DFF) splits deep features into sub-features and fuses them using a diffusion process to integrate complementary information.
• The DFF process models each sub-feature as a separate graph, then fuses metrics across graphs to generate a unified, improved distance metric.
• The ARR-DFF strategy applies ARR on sub-feature metrics and fuses the resulting metrics via DFF, enabling synergistic performance gains.
• The method is designed to be plug-and-play, compatible with existing re-ID pipelines using standard deep features and distance metrics.

Experimental results

Research questions

• RQ1 Can iterative re-ranking be made more effective by adaptively increasing the amount of contextual information used in each iteration?
• RQ2 Can the diversity of sub-features within a deep feature representation be exploited to improve re-ID performance?
• RQ3 Does combining adaptive re-ranking with deep feature fusion lead to superior performance compared to individual components?
• RQ4 How sensitive are the proposed methods to hyperparameters such as the number of sub-features and the number of nearest neighbors?

Key findings

• ARR-DFF achieves a mAP of 87.6% and R-1 accuracy of 91.5% on the Market-1501 dataset, outperforming SPreID by 4.2 percentage points in mAP.
• On DukeMTMC-reID, ARR-DFF achieves 80.8% mAP and 85.0% R-1, surpassing SPreID by 7.5 percentage points in mAP.
• On CUHK03, ARR-DFF achieves 74.2% mAP and 71.6% R-1, demonstrating strong performance on a challenging, noisy dataset.
• The ARR-DFF strategy consistently outperforms DFF-ARR, with a 4.2 percentage point gain in R-1 accuracy and a 2.8 percentage point gain in mAP on Market-1501.
• The method is robust to hyperparameter choices, with performance stable across different values of S (number of sub-features) and k (number of nearest neighbors).
• DFF-ARR underperforms compared to ARR-DFF, indicating that the order of operations matters, with re-ranking before fusion being more effective.