Graph-Based Contrastive Learning for Description and Detection of Local Features.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• State-of-the-art local feature detection methods struggle with repetitive textures due to "pseudo-negatives."
• This leads to inconsistent optimization objectives during model training.

Purpose of the Study:

• To develop a self-supervised graph-based contrastive learning framework, GCLFeat, for robust local feature representation.
• To specifically address and alleviate the issue of "pseudo-negatives" in challenging environments.

Main Methods:

• Designed a graph neural network (GNN) to mine local transformational invariance and global textual knowledge.
• Generated dense correspondence annotations using a self-supervised paradigm on diverse datasets.
• Implemented a keypoints-aware sampling strategy for effective loss computation.

Main Results:

• The unsupervised GCLFeat framework demonstrated superior performance compared to state-of-the-art supervised baselines.
• Significant improvements were observed across various downstream tasks, including image matching, 3-D reconstruction, and visual localization.

Conclusions:

• GCLFeat offers an effective unsupervised approach for learning local features, particularly in environments with repetitive textures.
• The framework's ability to handle "pseudo-negatives" enhances its applicability in real-world computer vision challenges.