A Vectorial Approach to Unbalanced Optimal Mass Transport

Jiening Zhu, Rena Elkin, Jung Hun Oh, Joseph O. Deasy, Allen R. TannenbaumIEEE Access 2020 · DOI

TL;DR

This paper presents a novel way to unify unbalanced optimal mass transport (OMT) and vector-valued OMT by reformulating unbalanced OMT as a special case of vector-valued OMT. This unification simplifies implementation by allowing existing vector-valued OMT codes to handle unbalanced scenarios with minimal modifications, primarily by introducing a "source layer" and adjusting weighting parameters.

What Problem Does This Paper Address?

Optimal Mass Transport (OMT) is a powerful mathematical framework for moving mass from one distribution to another, with applications spanning machine learning, computer vision, and physics. Standard OMT assumes mass is conserved during transport. However, many real-world scenarios involve mass creation or destruction, which is addressed by "unbalanced OMT." Separately, OMT has also been extended to handle vector-valued distributions (e.g., color images), where mass can flow not only spatially but also between different channels. This paper tackles the challenge of unifying these two extensions: how to elegantly incorporate unbalanced mass transport within the existing framework of vector-valued OMT, making it easier to implement and potentially leading to new insights.

Methods

The core contribution of this paper is a reformulation that embeds unbalanced OMT into the vector-valued OMT framework. This is achieved by introducing an additional "source layer" that acts as a conduit for mass creation or destruction.

  • Unbalanced OMT as a Vector-Valued Problem: The authors propose that unbalanced OMT can be viewed as a vector-valued OMT problem with an extra layer. This layer represents the source/sink of mass. The continuity equation for unbalanced OMT, which includes a source term ($s$), is reinterpreted in the vector-valued context where the flow into the source layer acts as this source term.
  • "Source Layer" Concept: For scalar unbalanced OMT, a new layer is added parallel to the original space. An edge connects each point in the original space to this source layer. The flow on this edge is used to model the mass creation or destruction.
  • Weighting Parameters: To make the unbalanced OMT formulation equivalent to the vector-valued OMT with a source layer, specific weighting parameters are introduced. By assigning very small weights to the source layer and its associated flows, the energy contribution from this layer becomes negligible, effectively mimicking the behavior of unbalanced OMT.
  • Vector-Valued Unbalanced OMT: The approach is extended to vector-valued unbalanced OMT by adding a source layer connected to each of the existing vector channels. The difference in total mass between the source and target distributions is placed in this new layer, preserving the total mass of the augmented (n+1)-vector structure.

Datasets

The paper uses numerical examples to demonstrate its formulation. These include:

  • Moving two Gaussian distributions (for scalar unbalanced OMT).
  • Testing with images exhibiting very large total mass differences.
  • Color image data (for vector-valued unbalanced OMT).

The numerical algorithm used for these tests is based on [7]. No new datasets were introduced.

Key Insights and Ideas

  • Unification is Key: The most significant insight is that unbalanced OMT can be elegantly framed as a special instance of vector-valued OMT. This unification is not just theoretical but has practical implications for implementation.
  • The "Source Layer" Analogy: The introduction of a "source layer" provides an intuitive way to visualize and mathematically model mass creation and destruction within the OMT framework. This layer acts as a flexible sink or source for mass.
  • Weighting as a Control Mechanism: The paper highlights how carefully chosen weighting parameters can effectively "turn on" or "turn off" the influence of the source layer, allowing the unified model to seamlessly transition between balanced and unbalanced transport scenarios.
  • Simplified Implementation: By reformulating unbalanced OMT into a vector-valued OMT problem, existing, potentially complex, vector-valued OMT codes can be adapted with minimal changes (adding a layer and adjusting weights) to handle unbalanced cases.

Key Results

The paper demonstrates the effectiveness of its reformulation through numerical examples.

  • Scalar Unbalanced OMT: Figures 4 and 5 show how tuning the weight parameter $\gamma$ influences the transport. With large $\gamma$, the source layer is almost zero, indicating mass conservation. With small $\gamma$, mass flows through the source layer, demonstrating its role in handling mass differences.
  • Vector-Valued Unbalanced OMT: Figures 9 and 10 illustrate the application to color images, showing how the vector-valued distributions evolve over time, with the source layer capturing mass differences.
  • Qualitative Success: The numerical results qualitatively show that the proposed method can handle both scenarios where mass is conserved and where it is not, and that the weighting parameters effectively control the degree of imbalance.

Limitations and Future Work

  • Limited Unbalanced Formulation: The authors acknowledge that they only considered one specific type of unbalanced formulation (adding a source term). Other flow-based unbalanced OMT settings might also fit their model but were not explored.
  • Weighting Parameter Tuning: While the paper shows the effect of weighting parameters, the optimal selection of these parameters for specific applications is not detailed.
  • Future Research Directions: The authors suggest that their generalization of vector-valued OMT has not reached its full potential. Future work could involve:
  • Investigating different weighting schemes: applying different weights not just to layers, but also to specific areas, nodes, or edges within the graph structure.
  • Exploring other unbalanced OMT formulations within this unified framework.
  • Developing more sophisticated methods for automatically determining optimal weighting parameters based on application needs.

Summary updated

BibTeX
@article{zhu2020vectorial,
  author = {Jiening Zhu and Rena Elkin and Jung Hun Oh and Joseph O. Deasy and Allen R. Tannenbaum},
  title = {{A Vectorial Approach to Unbalanced Optimal Mass Transport}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.3038567},
}
Figure