From Keypoints to Object Landmarks via Self-Training Correspondence: A Novel Approach to Unsupervised Landmark...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised learning of object landmark detectors is challenging.
• Existing methods often rely on auxiliary tasks like image generation or equivariance.
• A need exists for more robust and flexible landmark detection methods.

Purpose of the Study:

• To propose a novel self-training paradigm for unsupervised learning of object landmark detectors.
• To develop an iterative algorithm that refines generic keypoints into distinctive landmarks.
• To enhance the flexibility of landmark detectors in capturing large viewpoint changes.

Main Methods:

• A self-training approach is employed, diverging from auxiliary tasks.
• An iterative algorithm alternates between feature clustering for pseudo-label generation and contrastive learning for feature distinction.
• A shared backbone is utilized for both the landmark detector and descriptor.

Main Results:

• The keypoint locations progressively converge to stable landmarks by filtering unstable ones.
• The method achieves new state-of-the-art results on diverse datasets (LS3D, BBCPose, Human3.6M, PennAction).
• The learned landmarks demonstrate greater flexibility in handling significant viewpoint variations.

Conclusions:

• The proposed self-training method offers a novel and effective approach to unsupervised landmark detection.
• The iterative refinement process successfully transforms keypoints into stable and distinctive landmarks.
• The method sets a new benchmark for performance on challenging object landmark detection tasks.