Learning Local Feature Descriptors Using Convex Optimisation.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Sparse feature detectors are crucial for viewpoint-invariant matching in computer vision.
• Existing methods for learning visual descriptors have limitations in handling sparse features and large datasets.

Purpose of the Study:

• To develop novel, discriminative learning methods for visual descriptors tailored to sparse feature detectors.
• To improve the performance of viewpoint-invariant matching using learned descriptors.
• To extend descriptor learning to weakly supervised scenarios using unannotated image collections.

Main Methods:

• Formulating descriptor pooling region learning as a sparse convex optimization problem.
• Applying Mahalanobis matrix nuclear norm regularization for descriptor dimensionality reduction via convex optimization.
• Utilizing discriminative large margin learning constraints for descriptor optimization.
• Evaluating descriptor performance after binarization for compression.
• Developing weakly supervised learning formulations for descriptor learning from unannotated data.

Main Results:

• The proposed convex optimization methods effectively learn descriptor pooling regions and reduce dimensionality.
• Learned descriptors, when binarized, show competitive performance.
• The weakly supervised approach enables descriptor learning from large, unannotated image collections.
• The new learning methods outperform the state-of-the-art on benchmark datasets (Brown et al., Philbin et al.).

Conclusions:

• This work presents a significant advancement in learning robust visual descriptors for sparse feature detection.
• The developed methods offer improved performance and flexibility, particularly in weakly supervised settings.
• The findings contribute to more efficient and accurate image matching and retrieval systems.