[Paper Review] Deep Learning-based Occluded Person Re-identification: A Survey

Yunjie Peng, Saihui Hou+4 more|arXiv (Cornell University)|Jul 29, 2022

Video Surveillance and Tracking Methods4 citations

TL;DR

This survey provides a comprehensive analysis of deep learning-based occluded person re-identification (Re-ID), categorizing methods by four core issues: position misalignment, scale misalignment, noisy information, and missing information. It evaluates recent models on four benchmark datasets, revealing that attention mechanisms and auxiliary supervision significantly improve performance, with top methods achieving 86.3% mAP on Occ-ReID and 87.3% mAP on Occ-DukeMTMC.

ABSTRACT

Occluded person re-identification (Re-ID) aims at addressing the occlusion problem when retrieving the person of interest across multiple cameras. With the promotion of deep learning technology and the increasing demand for intelligent video surveillance, the frequent occlusion in real-world applications has made occluded person Re-ID draw considerable interest from researchers. A large number of occluded person Re-ID methods have been proposed while there are few surveys that focus on occlusion. To fill this gap and help boost future research, this paper provides a systematic survey of occluded person Re-ID. Through an in-depth analysis of the occlusion in person Re-ID, most existing methods are found to only consider part of the problems brought by occlusion. Therefore, we review occlusion-related person Re-ID methods from the perspective of issues and solutions. We summarize four issues caused by occlusion in person Re-ID, i.e., position misalignment, scale misalignment, noisy information, and missing information. The occlusion-related methods addressing different issues are then categorized and introduced accordingly. After that, we summarize and compare the performance of recent occluded person Re-ID methods on four popular datasets: Partial-ReID, Partial-iLIDS, Occluded-ReID, and Occluded-DukeMTMC. Finally, we provide insights on promising future research directions.

Motivation & Objective

• To systematically categorize and analyze deep learning-based methods for occluded person Re-ID, focusing on the specific challenges introduced by occlusion.
• To identify and classify existing approaches according to the four primary issues caused by occlusion: position misalignment, scale misalignment, noisy information, and missing information.
• To compare the performance of recent state-of-the-art methods on four widely used occluded Re-ID benchmarks: Partial-ReID, Partial-iLIDS, Occluded-ReID, and Occluded-DukeMTMC.
• To provide insights into promising future research directions in occluded person Re-ID based on a critical review of current limitations and trends.

Proposed method

• The paper organizes existing methods into a multi-dimensional taxonomy based on the four occlusion-induced issues: position misalignment, scale misalignment, noisy information, and missing information.
• It reviews technical solutions under each issue category, including matching-based methods, auxiliary models, attention mechanisms (e.g., self-attention/Transformer), and image transformation techniques.
• The survey evaluates methods using performance metrics such as Rank-1 accuracy and mean average precision (mAP) on four standard datasets: Partial-ReID, Partial-iLIDS, Occluded-ReID, and Occluded-DukeMTMC.
• It analyzes the integration of attention mechanisms (e.g., MHSA-Net, PAT, FED) and additional supervision (e.g., HPNet, PFD) to address position and noise issues.
• The paper compares methods that use spatial or temporal recovery for missing information, such as RFCNet’s dual-path recovery mechanism.
• It evaluates image transformation techniques that aim to align partial and holistic images by learning conditional transformations, though these show higher computational cost and lower performance.

Experimental results

Research questions

• RQ1 What are the four primary issues introduced by occlusion in person Re-ID, and how do they affect model performance?
• RQ2 How do existing deep learning-based methods address position misalignment and scale misalignment in occluded Re-ID?
• RQ3 What role do attention mechanisms and auxiliary supervision play in mitigating noisy and missing information in occluded Re-ID?
• RQ4 How do different technical routes (e.g., matching, attention, image transformation) compare in terms of performance across standard benchmarks?
• RQ5 What are the most promising future research directions in occluded person Re-ID based on current methodological limitations and performance trends?

Key findings

• The top-performing method on the Occ-DukeMTMC dataset, HPNet, achieved 87.3% mAP using additional supervision, significantly outperforming earlier methods.
• On the Occ-ReID dataset, the FED model achieved 86.3% mAP by leveraging attention mechanisms and additional supervision to handle noisy and missing information.
• Methods using attention mechanisms such as PAT and FED achieved over 80% mAP on Occ-ReID, demonstrating the effectiveness of self-attention in modeling long-range dependencies in partial images.
• Image transformation techniques, while theoretically promising for aligning partial and holistic images, showed lower performance and higher computational cost compared to other approaches.
• Auxiliary models and additional supervision (e.g., in HPNet, PFD) consistently improved performance on both Occ-DukeMTMC and Occ-ReID, especially for position and noise-related issues.
• The integration of spatial and temporal recovery mechanisms in RFCNet achieved 63.9% Rank-1 and 54.5% mAP on Occ-DukeMTMC, showing the value of multi-modal information recovery.