Progressive Learning for Person Re-Identification with One Example.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Person re-identification (re-ID) is challenging with limited labeled data.
• Existing methods struggle to effectively leverage abundant unlabeled data in the one-example re-ID setting.

Purpose of the Study:

• To develop a progressive framework for one-example person re-identification.
• To enhance Convolutional Neural Network (CNN) model performance by incorporating unlabeled data.
• To improve the efficiency of pseudo-labeling strategies in re-ID tasks.

Main Methods:

• A progressive framework iteratively updates CNN models and estimates pseudo-labels for unlabeled data.
• Joint training utilizes labeled, pseudo-labeled, and index-labeled data for robust model optimization.
• A progressive sampling strategy dynamically selects reliable pseudo-labeled candidates, gradually increasing their number.

Main Results:

• Achieved state-of-the-art rank-1 accuracy, outperforming existing methods by 21.6 points on MARS and 16.6 points on DukeMTMC-VideoReID.
• Demonstrated comparable performance to fully supervised methods even with only 20% labeled data in a few-example setting.

Conclusions:

• The proposed progressive framework effectively exploits unlabeled data for person re-ID.
• The novel joint training and progressive sampling strategies lead to significant performance gains.
• This approach offers a promising solution for data-scarce person re-identification scenarios.