[Paper Review] Diversity-Achieving Slow-DropBlock Network for Person Re-Identification

Xiaofu Wu, Ben Xie+4 more|arXiv (Cornell University)|Feb 9, 2020

Video Surveillance and Tracking Methods24 references4 citations

TL;DR

This paper proposes the Diversity-Achieving Slow-DropBlock (SDB) network for person re-identification by moving feature dropping from intermediate layers to the input image level, enabling greater feature diversity through multiple dropping branches with individual dropping ratios. A novel double-batch-split co-training strategy ensures training convergence, achieving state-of-the-art performance on Market-1501, DukeMTMC-reID, and CUHK03 with mAP gains up to 4.3% over prior methods.

ABSTRACT

A big challenge of person re-identification (Re-ID) using a multi-branch network architecture is to learn diverse features from the ID-labeled dataset. The 2-branch Batch DropBlock (BDB) network was recently proposed for achieving diversity between the global branch and the feature-dropping branch. In this paper, we propose to move the dropping operation from the intermediate feature layer towards the input (image dropping). Since it may drop a large portion of input images, this makes the training hard to converge. Hence, we propose a novel double-batch-split co-training approach for remedying this problem. In particular, we show that the feature diversity can be well achieved with the use of multiple dropping branches by setting individual dropping ratio for each branch. Empirical evidence demonstrates that the proposed method performs superior to BDB on popular person Re-ID datasets, including Market-1501, DukeMTMC-reID and CUHK03 and the use of more dropping branches can further boost the performance.

Motivation & Objective

• To address the challenge of learning diverse features in multi-branch person Re-ID networks using ID-labeled data.
• To overcome the training instability caused by aggressive input-level image dropping, which may hinder convergence.
• To extend the two-branch Batch DropBlock (BDB) framework to an arbitrary number of dropping branches for enhanced feature diversity.
• To develop a co-training strategy that enables stable training while maintaining high feature diversity across multiple dropping branches.
• To empirically validate the effectiveness of SDB across multiple standard Re-ID benchmarks.

Proposed method

• Proposes Slow-DropBlock (SDB), relocating the feature dropping operation from intermediate layers to the input image level to increase diversity in learned representations.
• Introduces a novel double-batch-split co-training approach that decouples the training of the global and local (dropping) branches, ensuring convergence despite aggressive input-level dropping.
• Employs multiple dropping branches with individually configurable dropping ratios (e.g., r_h = 0.1, 0.3, 0.5) to promote diverse feature learning across different spatial regions.
• Incorporates attention modules to enhance feature representation quality, particularly in challenging scenarios.
• Uses a dimension-reduction layer to improve performance on CUHK03, especially in the labeled split, where it boosts rank-1 accuracy by 10.0%.
• Employs a flexible rate-of-changes parameter Q to control how often the dropping pattern is updated during training, optimizing for convergence and performance.

Experimental results

Research questions

• RQ1 Can moving the dropping operation from intermediate layers to the input image level significantly improve feature diversity in person Re-ID?
• RQ2 How can stable training be achieved when using aggressive input-level image dropping that risks non-convergence?
• RQ3 Does extending the two-branch BDB framework to L-branch architectures (L ≥ 2) lead to consistent performance gains?
• RQ4 What is the optimal configuration of dropping ratios and training strategies (e.g., co-training) for maximizing performance across multiple Re-ID datasets?
• RQ5 How do attention modules and dimension-reduction layers affect performance, particularly on the CUHK03 dataset?

Key findings

• SDB-2 outperforms the two-branch BDB network by 4.3% in mAP on Market-1501, demonstrating the superiority of input-level dropping with multiple branches.
• The use of double-batch-split co-training improves mAP by approximately 3.0% on CUHK03 compared to single-batch training, confirming its effectiveness in stabilizing training.
• The optimal dropping ratio for SDB-2 on DukeMTMC-reID is r_h = 0.3, achieving 78.9% mAP and 89.8% rank-1 accuracy.
• Co-training consistently improves performance across all datasets, with the largest gain observed on CUHK03, indicating mutual reinforcement between global and local branches.
• Attention modules improve mAP by about 1% across all datasets, showing their value in enhancing feature discrimination.
• The dimension-reduction layer boosts rank-1 accuracy on CUHK03-Labeled by 10.0% (from 70.4% to 80.4%), highlighting its importance for this specific dataset.