Efficient Person Search: An Anchor-Free Approach (2109.00211v1)

Abstract: Person search aims to simultaneously localize and identify a query person from realistic, uncropped images. To achieve this goal, state-of-the-art models typically add a re-id branch upon two-stage detectors like Faster R-CNN. Owing to the ROI-Align operation, this pipeline yields promising accuracy as re-id features are explicitly aligned with the corresponding object regions, but in the meantime, it introduces high computational overhead due to dense object anchors. In this work, we present an anchor-free approach to efficiently tackling this challenging task, by introducing the following dedicated designs. First, we select an anchor-free detector (i.e., FCOS) as the prototype of our framework. Due to the lack of dense object anchors, it exhibits significantly higher efficiency compared with existing person search models. Second, when directly accommodating this anchor-free detector for person search, there exist several major challenges in learning robust re-id features, which we summarize as the misalignment issues in different levels (i.e., scale, region, and task). To address these issues, we propose an aligned feature aggregation module to generate more discriminative and robust feature embeddings. Accordingly, we name our model as Feature-Aligned Person Search Network (AlignPS). Third, by investigating the advantages of both anchor-based and anchor-free models, we further augment AlignPS with an ROI-Align head, which significantly improves the robustness of re-id features while still keeping our model highly efficient. Extensive experiments conducted on two challenging benchmarks (i.e., CUHK-SYSU and PRW) demonstrate that our framework achieves state-of-the-art or competitive performance, while displaying higher efficiency. All the source codes, data, and trained models are available at: https://github.com/daodaofr/alignps.

• The paper introduces AlignPS, an anchor-free framework that eliminates dense region proposals for efficient person search.
• It integrates an Aligned Feature Aggregation module and a TOIM loss to robustly align multi-scale features and enhance re-identification.
• Experimental results on CUHK-SYSU and PRW benchmarks show competitive mAP and top-1 accuracy at faster speeds.

Efficient Person Search: An Anchor-Free Approach

The paper "Efficient Person Search: An Anchor-Free Approach" introduces a novel framework for person search, addressing the need for efficient identification and localization of individuals in complex scene images. The primary innovation in this research is the implementation of an anchor-free model, named Feature-Aligned Person Search Network (AlignPS), which fundamentally rethinks the architectures traditionally used in person search tasks.

Background and Motivation

Person search combines the challenges of pedestrian detection and person re-identification (re-id), typically handled by complex models with high computational loads. Existing state-of-the-art techniques often rely on Faster R-CNN with a re-id branch, using ROI-Align for feature alignment, leading to substantial computational demands due to the utilization of dense object anchors. This paper seeks to enhance computational efficiency while maintaining or improving the accuracy of person search algorithms.

Methodology

The central contribution of this paper is the introduction of AlignPS, an anchor-free framework derived from the Fully Convolutional One-Stage object detector (FCOS). Noteworthy improvements and designs in AlignPS include:

1. Anchor-Free Architecture: By leveraging FCOS, AlignPS eliminates dense region proposals and anchors, significantly enhancing efficiency.
2. Aligned Feature Aggregation (AFA) Module: This module addresses scale, region, and task misalignments by:
   • Using a single FPN (Feature Pyramid Network) level to handle scale variations, improving feature consistency across varying object sizes.
   • Incorporating deformable convolutions to dynamically adjust receptive fields, aligning features more robustly with person regions.
   • Prioritizing re-id learning by structuring the network to give precedence to re-id tasks, addressing the dual objectives within person search.
3. Improved Loss Function: The paper proposes a Triplet-aided Online Instance Matching (TOIM) loss, incorporating both intra-batch sample relationships and external identity latents for more discriminative learning.
4. ROI-AlignPS Variant: By adding a lightweight ROI-Align head during training, this variant enhances region alignment explicitly. During inference, it eschews RPN, thus maintaining high efficiency while improving identity features via a novel branch-level feature mutual learning strategy.

Results and Implications

The authors conducted extensive experiments on the CUHK-SYSU and PRW benchmarks. AlignPS delivered competitive or superior performance compared to contemporary methods while operating at faster speeds. Particularly, ROI-AlignPS showed a remarkable balance between computational efficiency and re-id accuracy, outperforming established methods like NAE+ and SeqNet in mean average precision (mAP) and top-1 accuracy.

Discussions and Future Work

The research delineates crucial insights into the advantages of anchor-free models within person search frameworks, offering significant implications for real-world surveillance and identification systems. Given the lightweight nature and increased speed, frameworks like AlignPS are invaluable in applications requiring rapid processing and deployment on devices with constrained resources. Future work could explore further imperfections such as the handling of extremely small object detections and scaling the architecture to broader datasets.

In summary, this paper provides a comprehensive analysis and substantial improvements on the person search task through a sophisticated blend of anchor-free design, feature alignment, and loss optimization, setting a new precedent for future advancements in the domain.