[Paper Review] Speeding-up Graphical Model Optimization via a Coarse-to-fine Cascade of Pruning Classifiers

Bruno Conejo, Nikos Komodakis+2 more|arXiv (Cornell University)|Sep 15, 2014

Advanced Image and Video Retrieval Techniques31 references3 citations

TL;DR

This paper proposes a coarse-to-fine cascade of learned pruning classifiers that dramatically accelerates Markov Random Field (MRF) optimization in computer vision tasks while improving solution accuracy. By adaptively pruning inactive labels at each scale using trained classifiers, the method achieves up to 4x speedup over state-of-the-art inference with lower energy and over 99% label agreement.

ABSTRACT

We propose a general and versatile framework that significantly speeds-up graphical model optimization while maintaining an excellent solution accuracy. The proposed approach relies on a multi-scale pruning scheme that is able to progressively reduce the solution space by use of a novel strategy based on a coarse-to-fine cascade of learnt classifiers. We thoroughly experiment with classic computer vision related MRF problems, where our framework constantly yields a significant time speed-up (with respect to the most efficient inference methods) and obtains a more accurate solution than directly optimizing the MRF.

Motivation & Objective

• Address the high computational cost of MRF optimization in large-scale vision problems with super-linear complexity.
• Overcome limitations of existing pruning and multi-scale methods that rely on rigid heuristics or risk losing optimal solutions.
• Develop a general, learnable framework that progressively reduces the solution space without sacrificing accuracy.
• Enable significant speedups while maintaining or improving energy minimization performance across diverse vision tasks.

Proposed method

• Employ a coarse-to-fine multi-scale representation of the graphical model to progressively refine the MAP estimate.
• Introduce a cascade of trained binary classifiers to predict which labels can be safely pruned at each scale, per node.
• Use a fine-grained pruning strategy that allows arbitrary, node-specific label subsets to be discarded based on classifier confidence.
• Train the classifiers on features derived from the current scale’s solution and local image context to predict non-optimal labels.
• Apply pruning at multiple scales with increasing aggressiveness, controlled by a hyperparameter λ, to balance speed and accuracy.
• Integrate the pruning framework into existing MRF solvers (e.g., FastPD) without modifying the core optimization algorithm.

Experimental results

Research questions

• RQ1 Can a learnable, coarse-to-fine pruning cascade outperform traditional MRF optimization in terms of speed and accuracy?
• RQ2 How effective is classifier-based label pruning in preserving the optimal solution, especially near discontinuities?
• RQ3 To what extent can the solution space be reduced without degrading the energy minimization result?
• RQ4 Does the framework generalize across diverse MRF problems such as stereo matching, optical flow, and image restoration?
• RQ5 Can the method achieve better energy values than direct optimization, even with aggressive pruning?

Key findings

• The proposed pruning-based optimization achieves a speed-up factor of up to 4x compared to FastPD, the most efficient baseline method.
• For all tested problems, the method consistently obtains lower energy values than direct optimization, indicating improved solution quality.
• Label agreement exceeds 99% in most cases, with the best agreement never dropping below 96%, even under aggressive pruning.
• The active label ratio correlates strongly with speed-up, confirming that pruning effectively reduces the search space.
• The framework preserves accuracy near label discontinuities, as evidenced by reduced pruning in such regions across all datasets.
• The method generalizes well: trained classifiers on one dataset (e.g., Tsukuba) successfully prune labels on unseen datasets (e.g., Venus, Teddy, House).