Progressive Transfer Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Model fine-tuning is crucial for transfer learning in person Re-identification (ReID).
• Mini-batch training in ReID faces challenges due to distribution gaps caused by intra-scenario variations (viewpoint, illumination, occlusion).

Purpose of the Study:

• To address the challenge of utilizing global dataset information during mini-batch training for model fine-tuning.
• To improve the performance of ReID models in diverse and challenging target scenarios.

Main Methods:

• Introduced Batch-related Convolutional Cell (BConv-Cell) to progressively aggregate global dataset information into a latent state.
• Proposed the Progressive Transfer Learning (PTL) method, jointly optimizing BConv-Cells and pre-trained ReID models for enhanced fine-tuning.
• Extended the PTL method to general image classification tasks.

Main Results:

• Significantly improved ReID model performance on benchmark datasets: MSMT17, Market-1501, CUHK03, and DukeMTMC-reID.
• Demonstrated substantial performance gains for baseline models in general image classification tasks.
• Empirical experiments validated the effectiveness of the proposed BConv-Cell and PTL method.

Conclusions:

• The proposed BConv-Cell and PTL method effectively bridge the distribution gap in mini-batch training by leveraging global dataset information.
• The approach offers a robust solution for enhancing transfer learning in person ReID and generalizes well to image classification.
• The method provides a significant advancement in fine-tuning deep learning models for visual recognition tasks.