Pose Invariant Embedding for Deep Person Re-identification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Pedestrian misalignment due to detector errors and pose variations hinders robust person re-identification (re-ID).
• Poor alignment compromises feature learning and matching, impacting system performance.

Purpose of the Study:

• To introduce a pose invariant embedding (PIE) as a novel pedestrian descriptor.
• To address the critical problem of pedestrian misalignment in person re-ID systems.

Main Methods:

• Introduced the PoseBox structure via pose estimation and affine transformations for standard pedestrian alignment.
• Designed a PoseBox fusion (PBF) CNN architecture integrating original images, PoseBox, and pose estimation confidence.
• Defined the PIE descriptor as the fully connected layer of the PBF network for retrieval tasks.

Main Results:

• PoseBox alone demonstrated decent re-ID accuracy.
• The PIE descriptor, learned via the PBF network, achieved competitive performance against state-of-the-art approaches.
• Experiments were validated on Market-1501, CUHK03-NP, and DukeMTMC-reID datasets.

Conclusions:

• The proposed PIE descriptor effectively mitigates issues caused by pedestrian misalignment.
• The PBF CNN architecture successfully reduces the impact of pose estimation errors and information loss.
• The developed method offers a significant advancement for robust person re-identification.