[Paper Review] A Coarse-to-fine Pyramidal Model for Person Re-identification via Multi-Loss Dynamic Training.

Feng Zheng, Xing Sun+4 more|arXiv (Cornell University)|Oct 29, 2018

Video Surveillance and Tracking Methods15 citations

TL;DR

This paper proposes a coarse-to-fine pyramidal model for person re-identification that leverages multi-scale features and a dynamic training scheme to reduce reliance on precise bounding boxes. By integrating local and global cues across scales and unifying two losses through dynamic training, the method achieves state-of-the-art performance, outperforming the previous best method by 9.5% on the challenging CUHK03 dataset.

ABSTRACT

Most existing Re-IDentification (Re-ID) methods are highly dependent on precise bounding boxes that enable images to be aligned with each other. However, due to the inevitable challenging scenarios, current detection models often output inaccurate bounding boxes yet, which inevitably worsen the performance of these Re-ID algorithms. In this paper, to relax the requirement, we propose a novel coarse-to-fine pyramid model that not only incorporates local and global information, but also integrates the gradual cues between them. The pyramid model is able to match the cues at different scales and then search for the correct image of the same identity even when the image pair are not aligned. In addition, in order to learn discriminative identity representation, we explore a dynamic training scheme to seamlessly unify two losses and extract appropriate shared information between them. Experimental results clearly demonstrate that the proposed method achieves the state-of-the-art results on three datasets and it is worth noting that our approach exceeds the current best method by 9.5% on the most challenging dataset CUHK03.

Motivation & Objective

• To reduce the dependency of person re-identification models on precise, ground-truth bounding boxes, which are often inaccurate in real-world scenarios.
• To improve re-identification performance under misaligned or poorly detected image pairs by leveraging multi-scale feature representations.
• To develop a dynamic training strategy that seamlessly unifies multiple loss functions for better feature learning.
• To integrate coarse and fine-grained features across a pyramid structure to enhance discriminative identity representation.

Proposed method

• The method employs a coarse-to-fine pyramidal architecture that captures features at multiple scales, enabling robust matching even when images are not precisely aligned.
• Local and global features are progressively fused across pyramid levels to enhance representation learning across varying levels of detail.
• A dynamic training scheme is introduced to unify two loss functions, adaptively balancing their contributions during optimization.
• The model learns discriminative identity features by jointly optimizing for identity classification and metric learning through the dynamic loss integration.
• The pyramid structure enables the model to search for correct matches across scales, improving robustness to misalignment.

Experimental results

Research questions

• RQ1 Can a multi-scale pyramid model improve person re-identification performance when bounding boxes are inaccurate?
• RQ2 How can multiple loss functions be effectively combined during training to enhance feature discriminability?
• RQ3 To what extent can coarse-to-fine feature fusion reduce reliance on precise object localization in Re-ID?
• RQ4 Does the proposed dynamic training scheme lead to better generalization and performance on challenging Re-ID benchmarks?

Key findings

• The proposed method achieves state-of-the-art performance on three benchmark person re-identification datasets.
• On the most challenging CUHK03 dataset, the method exceeds the previous best method by 9.5% in re-identification accuracy.
• The model demonstrates strong robustness to misaligned or poorly detected bounding boxes due to its multi-scale feature integration.
• The dynamic training scheme successfully unifies two loss functions, leading to improved feature learning and discriminative representation.