Related Experiment Video
Updated: Dec 28, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
Person Search by Separated Modeling and A Mask-Guided Two-Stream CNN Model
Summary
This study introduces a novel person search method by separating detection and re-identification (re-ID) feature extraction. This approach enhances identity representation and achieves state-of-the-art results on benchmark datasets.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Machine Learning
Background:
- Person search combines pedestrian detection and person re-identification (re-ID), which are complex tasks.
- Existing joint models often share representations, potentially limiting performance.
Purpose of the Study:
- To develop a more effective person search system by decoupling detection and re-ID feature extraction.
- To improve the representational power for individual identities.
Main Methods:
- Segmenting foreground persons from image patches to extract identity-specific features.
- Employing two separate Convolutional Neural Network (CNN) streams to model foreground persons and original image patches individually.
- Introducing Confidence Weighted Stream Attention to dynamically adjust stream importance based on detection confidence.
- Integrating semantic segmentation into the re-ID network, trained with weakly-annotated bounding boxes.
Main Results:
- Achieved state-of-the-art mean Average Precision (mAP) of 83.3% on the CUHK-SYSU benchmark.
- Achieved state-of-the-art mAP of 32.8% on the PRW benchmark.
- Demonstrated superior performance compared to existing methods through extensive ablation studies.
Conclusions:
- Separating detector and re-ID feature extraction leads to improved person search performance.
- The proposed method, incorporating stream attention and semantic segmentation, offers an effective and simplified pipeline.
- Results significantly surpass the current state of the art on standard benchmarks.
