PDC-Net+: Enhanced Probabilistic Dense Correspondence Network.

Area of Science:

• Computer Vision
• Machine Learning
• Geometric Deep Learning

Background:

• Dense correspondence estimation is crucial for computer vision tasks.
• Existing dense methods struggle with large displacements, occlusions, and homogeneous regions.
• Accurate confidence estimation is needed for reliable dense matching.

Purpose of the Study:

• To develop a network for accurate dense correspondences and reliable confidence maps.
• To address limitations of current dense flow estimation methods.
• To improve the applicability of dense methods in real-world scenarios.

Main Methods:

• Proposed the Enhanced Probabilistic Dense Correspondence Network (PDC-Net+).
• Developed a flexible probabilistic approach to jointly learn flow prediction and uncertainty.
• Utilized a constrained mixture model for predictive distribution and an enhanced self-supervised training strategy.

Main Results:

• Achieved state-of-the-art results on challenging geometric matching and optical flow datasets.
• Demonstrated accurate dense correspondence estimation with reliable confidence maps.
• Validated the effectiveness of probabilistic confidence estimation for downstream tasks.

Conclusions:

• PDC-Net+ provides robust and accurate dense correspondences with reliable uncertainty estimation.
• The proposed probabilistic approach and training strategy enhance generalization.
• The method significantly benefits applications like pose estimation, 3D reconstruction, and image retrieval.