[Paper Review] Deep Miner: A Deep and Multi-branch Network which Mines Rich and Diverse Features for Person Re-identification

Abdallah Benzine, Mohamed El Amine Seddik+1 more|arXiv (Cornell University)|Feb 18, 2021

Video Surveillance and Tracking Methods36 references4 citations

TL;DR

Deep Miner proposes a multi-branch CNN architecture that enhances person re-identification by mining rich, diverse features through three complementary branches: a global branch, a local part-based branch, and input-erased branches (IE-branches) that recover features ignored by the backbone. By applying a simple suppression operation to erase high-activation regions in feature maps, IE-branches force the network to learn new, discriminative features, achieving a 6.5% mAP gain over SOTA on MSMT17.

ABSTRACT

Most recent person re-identification approaches are based on the use of deep convolutional neural networks (CNNs). These networks, although effective in multiple tasks such as classification or object detection, tend to focus on the most discriminative part of an object rather than retrieving all its relevant features. This behavior penalizes the performance of a CNN for the re-identification task, since it should identify diverse and fine grained features. It is then essential to make the network learn a wide variety of finer characteristics in order to make the re-identification process of people effective and robust to finer changes. In this article, we introduce Deep Miner, a method that allows CNNs to "mine" richer and more diverse features about people for their re-identification. Deep Miner is specifically composed of three types of branches: a Global branch (G-branch), a Local branch (L-branch) and an Input-Erased branch (IE-branch). G-branch corresponds to the initial backbone which predicts global characteristics, while L-branch retrieves part level resolution features. The IE-branch for its part, receives partially suppressed feature maps as input thereby allowing the network to "mine" new features (those ignored by G-branch) as output. For this special purpose, a dedicated suppression procedure for identifying and removing features within a given CNN is introduced. This suppression procedure has the major benefit of being simple, while it produces a model that significantly outperforms state-of-the-art (SOTA) re-identification methods. Specifically, we conduct experiments on four standard person re-identification benchmarks and witness an absolute performance gain up to 6.5% mAP compared to SOTA.

Motivation & Objective

• To address the limitation of CNNs in person re-identification, where models tend to focus on only the most discriminative features, ignoring fine-grained and diverse attributes.
• To improve robustness to appearance variations such as pose, lighting, and viewpoint changes by encouraging the network to learn a broader range of features.
• To develop a simple yet effective method that enhances feature diversity without increasing model complexity or requiring complex architectures.
• To demonstrate that feature mining through input-erased branches significantly improves performance on standard benchmarks, especially under low-data regimes.

Proposed method

• The model integrates three branches: a global branch (G-branch) using the original CNN backbone to extract standard global features.
• A local branch (L-branch) applies uniform partitioning to extract part-level features, enhancing fine-grained representation learning.
• Input-Erased branches (IE-branches) receive feature maps with high-activation regions suppressed via a simple suppression procedure, forcing the network to discover previously ignored features.
• The suppression operation selectively removes features based on activation magnitude, creating a new input space for IE-branches to mine complementary features.
• Attention modules are applied in the G-branch and IE-branches to improve feature representation and reduce background noise.
• The final features from all branches are fused for re-identification, enabling richer and more diverse person embeddings.

Experimental results

Research questions

• RQ1 Can a multi-branch CNN architecture that includes input-erased branches significantly improve feature diversity and representation richness in person re-identification?
• RQ2 Does the suppression-based feature mining strategy in IE-branches lead to better performance than standard global or part-based feature learning?
• RQ3 How does the proposed method compare to SOTA approaches in terms of mAP and Rank-1 accuracy across diverse benchmarks with varying data complexity?
• RQ4 Can the method generalize effectively even under low-data or challenging conditions, such as in CUHK03 with limited samples?
• RQ5 Does the simplicity of the suppression operation enable strong performance without requiring complex architectural modifications?

Key findings

• Deep Miner achieves a 6.5% absolute mAP improvement over the best SOTA method on the MSMT17 dataset, the largest and most complex benchmark in person re-identification.
• On DukeMTMC-ReID, Deep Miner achieves 84.7% mAP (labeled) and 81.4% mAP (detected), outperforming SOTA methods including SCSN and Pyramid.
• On CUHK03, Deep Miner surpasses PLR OSNet by 4.2% mAP on the labeled split and by 3.2% on the detected split, demonstrating strong performance under low-data conditions.
• Despite using only ResNet-50 and a simple 4-stripe local branch, Deep Miner outperforms more complex models like Pyramid (which uses a pyramidal feature set) and MGN.
• The model achieves state-of-the-art mAP on all four standard benchmarks: Market1501, DukeMTMC-ReID, CUHK03, and MSMT17, with consistent gains across all datasets.
• Feature visualization confirms that IE-branches discover new, relevant features—such as clothing patterns, bags, or accessories—often missed by the global branch.