Overcoming Data Deficiency for Multi-Person Pose Estimation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Multi-person pose estimation (MPPE) models struggle with complex scenes and data limitations.
• Existing pose datasets have deficiencies in instance complexity and realistic scene representation.
• Traditional data augmentation is insufficient for current MPPE challenges.

Purpose of the Study:

• To identify and address key deficiencies in existing pose datasets for MPPE.
• To propose novel methods for enriching MPPE training data.
• To improve the robustness and generalizability of MPPE models.

Main Methods:

• Developed a model-agnostic full-view data generation (Full-DG) method to create more balanced pose complexity and realistic scenes.
• Introduced an adaptive category-aware loss (AC-loss) to mitigate pixel-level imbalances between keypoints and backgrounds.
• Evaluated the proposed methods on representative MPPE models (HigherHRNet, HRNet).

Main Results:

• Full-DG and AC-loss significantly improved MPPE accuracy.
• Performance gains of 1.0%-2.9% AP on the COCO benchmark were achieved.
• Gains of 1.0%-5.1% AP on the CrowdPose benchmark were observed.

Conclusions:

• The proposed Full-DG method and AC-loss effectively enhance MPPE training data.
• These methods improve the robustness and generalizability of MPPE models.
• The approach offers a practical solution for improving MPPE accuracy on challenging benchmarks.