Multi-Information Flow CNN and Attribute-Aided Reranking for Person Reidentification

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Person reidentification (re-id) is crucial for surveillance and security.
• Existing methods face challenges with occlusions, viewpoint changes, and limited training data.
• Integrating multi-modal information can improve re-id performance.

Purpose of the Study:

• To propose a novel Multi-Information Flow Convolutional Neural Network (MiF-CNN) for enhanced person reidentification.
• To develop an attribute recognition network to complement the re-id model.
• To introduce an attribute-aided reranking algorithm for improved re-id accuracy.

Main Methods:

• Designed MiF-CNN with multilayer convolutional structures for deeper networks and feature reuse.
• Developed a person attribute recognition network using Long Short-Term Memory (LSTM) and attention mechanisms.
• Implemented an attribute-aided reranking algorithm by fusing MiF-CNN and attribute recognition outputs.

Main Results:

• MiF-CNN demonstrated sufficient training on small-scale datasets.
• Achieved outstanding accuracy in person reidentification across VIPeR, CUHK01, and Market1501 datasets.
• Attribute-aided reranking significantly improved re-id accuracy, as confirmed by contrast experiments.

Conclusions:

• The proposed MiF-CNN model is effective for person reidentification, particularly with limited data.
• Attribute recognition and reranking provide a valuable enhancement to person re-id systems.
• The integrated approach offers a robust solution for accurate person reidentification.