Inception Convolution and Feature Fusion for Person Search

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Deep learning and hardware advancements are revolutionizing computer vision tasks like object detection and instance segmentation.
• Person search, an integrated task, faces challenges due to limitations in existing frameworks like Seq-Net, which rely on Faster R-CNN.
• Faster R-CNN's parallel structure can lead to suboptimal re-identification (re-ID) performance caused by low-resolution and overlooked features during pedestrian detection.

Purpose of the Study:

• To propose a novel person search methodology that overcomes the limitations of current frameworks.
• To enhance the accuracy and efficiency of pedestrian retrieval in computer vision systems.
• To develop a more robust and integrated approach for person search tasks.

Main Methods:

• Implemented an inception convolution module (ICM) to replace standard convolutions in ResNet-50, enabling dynamic channel distribution.
• Developed a feature fusion module (FFM) to integrate multi-level convolutional information for improved data extraction accuracy.
• Introduced a double-head module (DHM) for Bounding Box regression, combining global and fine-grained details to boost pedestrian retrieval precision.

Main Results:

• The proposed inception convolution and feature fusion module (IC-FFM) method demonstrated superior accuracy compared to the benchmark Seq-Net on the CHUK-SYSU and PRW datasets.
• The new methodology significantly improved the accuracy of pedestrian retrieval by effectively combining global and fine-grained information.
• Experimental results confirmed the enhanced performance of the proposed person search approach.

Conclusions:

• The proposed IC-FFM method offers a more accurate and effective solution for person search tasks compared to existing Seq-Net benchmarks.
• The methodology is simpler and readily integrable into current two-stage computer vision frameworks.
• This work contributes to the advancement of person search technology by improving feature extraction and bounding box regression accuracy.