Learning linear discriminant projections for dimensionality reduction of image descriptors

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Local image descriptors are crucial for image analysis tasks.
• Dimensionality reduction and improved discriminability are key challenges.
• Existing discriminant projection methods often require ground truth correspondence data.

Purpose of the Study:

• To introduce Linear Discriminant Projections (LDP) for dimensionality reduction and enhanced discriminability of local image descriptors.
• To analyze LDP properties and compare it with state-of-the-art methods.
• To demonstrate the applicability of LDP in scenarios lacking correspondence ground truth.

Main Methods:

• Linear Discriminant Projections (LDP) technique for feature extraction.
• Utilizing simulated image transformation data for training.
• Extensive experimental evaluation on standard datasets for image matching and categorization.

Main Results:

• LDP achieves significant dimensionality reduction of local descriptors (e.g., from 128 to 30 dimensions).
• Demonstrated performance increases in image matching and categorization tasks.
• Simulated training data provides comparable results to real data with ground truth.

Conclusions:

• LDP is an effective method for reducing dimensionality and improving the discriminability of local image descriptors.
• Simulated data enables LDP's application in unsupervised or weakly supervised learning scenarios.
• LDP advances the state-of-the-art in recognition performance with substantial dimensionality reduction.