Abstract
This paper introduces a novel steepest descent flow in Banach spaces, extending the concept of generalized gradient descent to the setting of Finsler metrics. By utilizing non-Hilbertian norms, the framework incorporates structural priors—specifically piecewise rigidity—into curve evolution. This approach allows for the deformation of articulated shapes without requiring prior knowledge of joint locations. We provide a convergence theorem for the proposed Finsler gradient descent and demonstrate its efficacy in curve matching tasks, particularly for articulated human and animal silhouettes.
1. Introduction
Curve evolution is a fundamental technique in computer vision, widely applied to image segmentation and shape registration. Traditionally, these evolutions are governed by gradients derived from Riemannian metrics, typically based on $L^2$ or Sobolev norms. While these Hilbertian structures are mathematically tractable, they often impose global smoothness constraints that fail to capture the physical reality of articulated objects. For instance, when a human moves, limbs remain rigid while deformation is localized at the joints. Standard smooth gradients often result in unnatural stretching, leading optimization processes into poor local minima.
This research addresses the challenge of defining a descent flow that inherently favors piecewise rigid motions—where segments of a curve move independently as rigid bodies—without necessitating pre-segmentation of the object's parts.
2. Finsler Gradient Descent in Banach Spaces
The authors move beyond the Riemannian framework by adopting Finsler geometry. In a Riemannian setting, the metric at each point is defined by a quadratic norm. In contrast, Finsler metrics allow for more general Banach norms, providing the flexibility to shape the "geometry" of the search space according to specific priors.
2.1 Theoretical Foundation
The paper defines a generalized gradient descent where each step is determined by solving a constrained convex optimization problem in a Banach space. A significant contribution of this work is the formal proof of a convergence theorem, ensuring that the Finsler steepest descent method converges to a stationary point of the energy functional. This theoretical grounding justifies the use of non-standard norms to navigate non-convex optimization landscapes where the space's geometry is critical for avoiding local minima.
3. Modeling Piecewise Rigidity
To promote structured deformations, the authors introduce a penalty function based on the Total Variation (TV) of the deformation’s normal component.
- Sparsity and Joints: The $L^1$-like nature of the TV norm promotes sparsity in the derivative of the transformation. In the context of curve evolution, this means that "bending" is concentrated at discrete points (the joints), while the intervening segments remain rigid.
- Automatic Discovery: A key advantage of this formulation is that it does not require manual labeling of joints. The optimization process "discovers" the optimal articulation points dynamically as the curve evolves toward the target shape.
- Piecewise Similarity: The model is further extended to accommodate "piecewise similarity" transformations. This allows segments to undergo scaling (shrinking or growing) in addition to rotation and translation, which is particularly useful for matching shapes with varying proportions.
4. Numerical Implementation
The descent steps are discretized using finite element methods. Because the Finsler gradient is defined via a non-quadratic norm, the update step cannot be computed as a simple linear projection. Instead, each iteration is formulated as a Second Order Cone Program (SOCP). While more computationally intensive than standard gradient steps, the SOCP framework allows the algorithm to effectively navigate the complex constraints imposed by the piecewise rigid prior.
5. Experimental Results and Insights
The method was evaluated on synthetic articulated shapes (such as "U" and "T" shapes) and real-world silhouettes of humans and animals.
5.1 Geometry as a Prior
The experiments confirm that the choice of metric acts as a powerful prior. By shifting from a Riemannian to a Finsler geometry, the "path of least resistance" for the optimization changes. This allows the flow to bypass the local minima that typically trap Sobolev-based methods, which often become "tangled" when attempting to match complex articulated poses.
5.2 Global Matching
By enforcing local rigidity, the algorithm prioritizes moving entire limbs into their correct global orientations before performing fine-grained adjustments. This hierarchical-like behavior emerges naturally from the Finsler flow and proves far more robust for articulated registration than global smooth deformations.
6. Conclusion and Future Work
The introduction of Finsler steepest descent provides a robust mathematical framework for incorporating structural priors into curve evolution. While the method demonstrates superior performance in matching articulated shapes, several areas for future research remain:
- Computational Efficiency: The requirement to solve an SOCP at each iteration is a bottleneck for real-time applications.
- Theoretical Limits: Proving the existence of a continuous "limit flow" as the step size approaches zero remains an open theoretical challenge.
- Topological Constraints: Future iterations could incorporate mechanisms to prevent self-intersection of the curve during evolution.
In summary, this work demonstrates that by carefully choosing the geometry of the optimization space, one can solve complex, non-convex shape matching problems that are otherwise intractable for standard gradient descent methods.
Summary updated