Human Co-Parsing Guided Alignment for Occluded Person Re-Identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Occluded person re-identification (ReID) is difficult due to background noise and incomplete foreground data.
• Existing human parsing-based ReID methods struggle with performance heavily dependent on the parsing model.
• Supervised methods require costly annotations, while unsupervised methods yield suboptimal segmentation.

Purpose of the Study:

• To develop a novel framework for occluded person ReID that utilizes existing dataset information for supervision.
• To improve the accuracy and efficiency of person ReID in the presence of occlusions.
• To reduce the reliance on expensive manual annotations for human parsing models.

Main Methods:

• Proposed the Human Co-parsing Guided Alignment (HCGA) framework, integrating a weakly supervised human co-parsing network into the ReID network.
• Employed three consistency criteria (global semantics, local space, background) for weak supervision of the co-parsing network.
• Utilized semantic information and deep features from the ReID network for guided alignment to obtain foreground and human part features.

Main Results:

• The HCGA framework demonstrated superior performance on both occluded and holistic person ReID datasets.
• Achieved 70.2% Rank-1 accuracy and 57.5% mAP on the Occluded-DukeMTMC dataset.
• Effectively extracts features for foreground and human parts, enhancing occluded person ReID.

Conclusions:

• The proposed method effectively uses pre-existing information as supervision for human parsing, eliminating the need for extra annotations.
• The HCGA framework offers a promising solution for occluded person ReID by leveraging weakly supervised co-parsing.
• The approach significantly advances the state-of-the-art in handling occluded person re-identification tasks.