Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
110 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
44 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Joint Detection and Identification Feature Learning for Person Search (1604.01850v3)

Published 7 Apr 2016 in cs.CV

Abstract: Existing person re-identification benchmarks and methods mainly focus on matching cropped pedestrian images between queries and candidates. However, it is different from real-world scenarios where the annotations of pedestrian bounding boxes are unavailable and the target person needs to be searched from a gallery of whole scene images. To close the gap, we propose a new deep learning framework for person search. Instead of breaking it down into two separate tasks---pedestrian detection and person re-identification, we jointly handle both aspects in a single convolutional neural network. An Online Instance Matching (OIM) loss function is proposed to train the network effectively, which is scalable to datasets with numerous identities. To validate our approach, we collect and annotate a large-scale benchmark dataset for person search. It contains 18,184 images, 8,432 identities, and 96,143 pedestrian bounding boxes. Experiments show that our framework outperforms other separate approaches, and the proposed OIM loss function converges much faster and better than the conventional Softmax loss.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (5)
  1. Tong Xiao (119 papers)
  2. Shuang Li (203 papers)
  3. Bochao Wang (5 papers)
  4. Liang Lin (318 papers)
  5. Xiaogang Wang (230 papers)
Citations (800)

Summary

Joint Detection and Identification Feature Learning for Person Search

The paper "Joint Detection and Identification Feature Learning for Person Search" by Tong Xiao, Shuang Li, Bochao Wang, Liang Lin, and Xiaogang Wang addresses a significant gap in the domain of person re-identification (re-id). Traditional approaches have predominantly focused on matching manually cropped pedestrian images. This methodology, while useful in controlled settings, falls short of addressing the complexities inherent in real-world scenarios. In practical applications, pedestrian bounding boxes are not pre-defined, necessitating the search for a target person in whole scene images. This paper proposes a unified deep learning framework that simultaneously handles detection and re-identification within a single Convolutional Neural Network (CNN), introducing several key innovations such as the Online Instance Matching (OIM) loss function.

The proposed framework deviates from conventional methods that treat pedestrian detection and person re-id as separate tasks. Instead, it integrates both functionalities into a cohesive model. Specifically, the framework is built upon the ResNet-50 architecture, where a pedestrian proposal network generates candidate bounding boxes and an identification network extracts features for comparison against the query person's features. The sharing of convolutional feature maps between the two networks enhances the efficiency of the model. This joint optimization allows the pedestrian proposal network to focus on recalling potential candidates, with the identification network refining these results by eliminating false positives and adjusting misalignments.

A notable contribution of this work is the Online Instance Matching (OIM) loss function. Traditional re-id training methods employ pairwise or triplet loss functions, which are computationally expensive and do not scale well with large datasets. The OIM loss, in contrast, maintains a lookup table of features for labeled identities and a circular queue for features from unlabeled instances. During training, distances are computed between mini-batch samples and all registered entries, enabling efficient and scalable feature learning.

The authors introduce and validate their framework on a newly collected and annotated benchmark dataset for person search, which includes 18,184 images, 8,432 identities, and 96,143 pedestrian bounding boxes. Experimental results demonstrate that their framework outperforms baseline approaches that treat detection and re-id separately. Additionally, the OIM loss function demonstrates superior convergence speed and performance compared to conventional Softmax loss.

Implications and Future Directions

The implications of this research are manifold. Practically, the framework provides a more effective and scalable solution for real-world person search applications, such as video surveillance and criminal identification. The joint optimization of detection and identification tasks leads to more robust performance in diverse and uncontrolled environments.

Theoretically, the OIM loss function offers a new perspective on scalable re-id training. Its non-parametric nature addresses the limitations of large Softmax classifiers, opening avenues for further exploration in loss functions tailored to large-scale identity matching problems. This development can potentially influence other subfields of computer vision where similar challenges of large-class classification and instance retrieval are prevalent.

Looking forward, several directions could extend the impact of this work. Enhancements could be made by incorporating advanced data augmentation techniques or exploring the integration of additional contextual information to aid identification. The scalability of the OIM loss can be further tested on even larger datasets, pushing the boundaries of re-id performance. Additionally, the framework's robustness could be evaluated in more diverse scenarios, including low-resolution images and occlusions, to ensure applicability across a broad spectrum of real-world conditions.

Furthermore, this approach lays a foundational methodology that could be adapted for other instance matching problems beyond person re-identification, such as object tracking or fine-grained object recognition. The community could draw significant inspiration from this paper, advancing integrated solutions that align more closely with practical application needs.

In conclusion, the paper presents a meticulously designed framework that bridges a critical gap between theoretical research and practical application in person search. The joint detection and identification learning and the pioneering OIM loss function are substantial contributions to the field, promising enhanced efficacy and scalability. The presented framework and dataset provide valuable resources for ongoing and future research, fostering continued advancements in person re-identification and related areas.

Youtube Logo Streamline Icon: https://streamlinehq.com