Learning Sample Specific Weights for Late Fusion.

Area of Science:

• Machine Learning
• Computer Vision
• Pattern Recognition

Background:

• Late fusion combines multiple classifiers to enhance recognition accuracy.
• Existing methods use fixed fusion weights, ignoring sample-dependent classifier performance.
• This limitation hinders optimal performance across diverse data subsets.

Purpose of the Study:

• To propose a novel sample specific late fusion (SSLF) method.
• To address the limitations of fixed fusion weights in multi-classifier systems.
• To improve recognition accuracy by learning sample-specific fusion weights.

Main Methods:

• Developed SSLF by framing late fusion as an information propagation process.
• Introduced two SSLF variants: Ranking SSLF (R-SSLF) and Infinite Push SSLF (I-SSLF).
• Utilized graph Laplacian, RankSVM, and infinite push constraints, solved with gradient projection and ADMM algorithms.
• Integrated AnchorGraph for scalability on large datasets.

Main Results:

• SSLF methods demonstrated superior performance on large-scale image and video datasets.
• The proposed methods effectively learn sample-specific fusion weights.
• Achieved improved accuracy by optimizing fusion for individual samples.

Conclusions:

• SSLF represents a significant advancement in late fusion techniques.
• The ability to learn sample-specific weights enhances recognition system adaptability.
• This work pioneers sample-specific weight learning for late fusion, offering scalability and improved accuracy.