Automatic Ensemble Diffusion for 3D Shape and Image Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Diffusion processes enhance visual retrieval systems.
• Similarity fusion is crucial as single measures are insufficient.
• Combining diffusion and similarity fusion is a key research area.

Purpose of the Study:

• To revisit and offer new insights into fusion with diffusion methods.
• To address the susceptibility of existing algorithms to noisy similarities.
• To propose a novel framework, Regularized Ensemble Diffusion (RED), for robust retrieval.

Main Methods:

• Revisiting existing fusion with diffusion techniques.
• Developing Regularized Ensemble Diffusion (RED) with automatic weight learning.
• Formulating RED as a convex optimization problem solvable via an iteration-based solver.

Main Results:

• RED effectively suppresses the negative impacts of noisy similarities.
• The proposed method maintains computational complexity similar to conventional diffusion processes.
• State-of-the-art performances were achieved on 3D shape and image retrieval datasets.

Conclusions:

• Regularized Ensemble Diffusion (RED) offers a robust approach to visual retrieval.
• The automatic weight learning paradigm significantly improves resilience to noisy data.
• The framework demonstrates superior performance across various benchmark datasets.