Not Every Patch is Needed: Towards a More Efficient and Effective Backbone for Video-based Person Re-identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Video-based person re-identification (ReID) typically relies on exhaustive feature extraction from all video frames.
• Conventional methods using Convolutional Neural Networks (CNNs) or Vision Transformers (ViT) can be computationally intensive.
• Similarities across frames in ReID videos, due to slight human movements, suggest potential for computational optimization.

Purpose of the Study:

• To propose a novel, efficient, and effective plug-and-play backbone for video-based person ReID.
• To reduce the computational cost associated with traditional ReID feature extraction methods.
• To maintain or improve accuracy while significantly decreasing computational demands.

Main Methods:

• Introduced a patch selection mechanism to extract features only from crucial, non-repetitive image patches.
• Developed a novel network structure to generate and utilize pseudo frame global context.
• Addressed incomplete views arising from sparse input data.

Main Results:

• Achieved significant reduction in computational cost: 74% compared to ViT-B and 28% compared to ResNet50.
• Maintained accuracy comparable to ViT-B.
• Demonstrated superior accuracy compared to ResNet50.
• The proposed backbone offers both high performance and low computational cost.

Conclusions:

• The proposed selective feature extraction and pseudo global context approach is effective for video-based person ReID.
• This method offers a computationally efficient alternative to conventional exhaustive feature extraction techniques.
• The backbone provides a strong balance between accuracy and computational cost for ReID tasks.