TwinsReID: Person re-identification based on twins transformer's multi-level features.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Traditional person re-identification models rely on Convolutional Neural Networks (CNNs) for feature extraction, which have limitations in receptive field size and computational cost.
• CNNs use convolution operations to reduce feature map size, leading to limited local receptive fields and high computational demands.

Purpose of the Study:

• To develop an end-to-end person re-identification model, twinsReID, that integrates feature information across levels using Transformer's self-attention mechanism.
• To overcome the limitations of CNNs by leveraging the global receptive field and computational efficiency of Transformers.

Main Methods:

• The proposed twinsReID model replaces CNNs with the Twins-SVT Transformer architecture.
• Features are extracted from two different stages, processed through convolution and global adaptive average pooling, and combined.
• Multiple feature vectors are generated and fed into Triplet Loss, Cross-Entropy Loss, and Center-Loss functions.

Main Results:

• The twinsReID model achieved a mAP (mean Average Precision) of 85.4% and rank1 accuracy of 93.7% on the Market-1501 dataset.
• After re-ranking, performance improved to 93.6% mAP and 94.9% rank1 accuracy.
• The model demonstrated a reduction in parameter count compared to traditional CNN-based models.

Conclusions:

• The twinsReID model effectively integrates multi-level features using Transformers, outperforming traditional CNN approaches in person re-identification.
• The Transformer-based approach offers advantages in receptive field coverage and computational efficiency.
• The study highlights the potential of Transformers for advanced computer vision tasks like person re-identification.