Siamese comparative transformer-based network for unsupervised landmark detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Landmark detection is crucial for various computer vision applications.
• Existing methods often focus on image reconstruction, neglecting semantic relationships between landmarks.
• This limitation hinders the achievement of semantic representations in landmark detection.

Purpose of the Study:

• To develop a novel network that strengthens semantic connections among detected landmarks.
• To improve the representation and encoding of landmarks by perceiving global feature relationships.

Main Methods:

• Introduced a Siamese comparative transformer-based network.
• Employed a Siamese contrastive regularizer to enhance connections between semantically similar landmarks.
• Integrated a lightweight direction-guided Transformer into the image pose encoder for global feature perception.

Main Results:

• The proposed method effectively strengthens semantic connections among landmarks.
• Improved representation and encoding of landmarks through enhanced global feature perception.
• Achieved competitive performance on CelebA, AFLW, and Cat Heads benchmarks against unsupervised and supervised methods.

Conclusions:

• The novel Siamese comparative transformer network successfully addresses the limitations of existing landmark detection algorithms.
• The method enhances semantic understanding and achieves state-of-the-art or competitive results in unsupervised landmark detection.
• This approach offers a promising direction for future research in semantic landmark detection.