Abstract
Instance segmentation remains a challenging task in computer vision, often requiring a compromise between computational efficiency and boundary accuracy. Traditional contour-based methods struggle with a "fixed-point" limitation, where too few points fail to capture complex shapes and too many points lead to excessive computational overhead. This paper introduces VeinMask, a single-shot instance segmentation framework inspired by the hierarchical growth of leaf veins. By utilizing a dynamic growth model consisting of "major" and "minor" veins, the framework achieves state-of-the-art accuracy with significantly lower design complexity than previous methods. Our results on the COCO and SBD datasets demonstrate that VeinMask effectively balances detail and speed, providing a scalable solution for representing irregular object shapes.
Introduction
In the field of computer vision, instance segmentation—the task of identifying and outlining every individual object in an image—is vital for applications ranging from autonomous driving to medical imaging. Current methodologies generally fall into two categories: mask-based (pixel-wise) and contour-based (point-wise). Contour-based methods are often preferred for their speed, yet they face a persistent "Goldilocks" problem: determining the optimal number of points to represent a shape.
Standard models treat all edges with equal priority, leading to redundant calculations on smooth boundaries and insufficient detail on complex, jagged ones. VeinMask addresses this by mimicking the natural growth patterns of leaves, where a primary structure (major veins) provides the general form and secondary structures (minor veins) emerge only where additional detail is required.
Methodology
1. Vein Growth Simulation
VeinMask represents object masks as a directed graph rather than a flat set of coordinates. The process begins at the object's geometric centroid (the "root").
- Major Veins: These sprout from the root to establish a coarse boundary.
- Minor Veins: If a segment of the boundary is identified as complex or "twisty," the model identifies a node on the major vein and grows minor veins to refine the local detail.
This hierarchical approach allows for dynamic point allocation, focusing computational resources on intricate areas while maintaining simplicity elsewhere.
2. Centroidness
Most instance segmentation models use a "centerness" metric to rank the quality of a predicted bounding box. However, for irregular or elongated shapes, the center of a bounding box rarely aligns with the actual center of the object. VeinMask introduces Centroidness, which utilizes a Gaussian distribution centered on the object's true geometric centroid. This shift reduces information distortion and allows the model to more effectively suppress low-quality detections during inference.
3. SCCS and R-IoU Loss
To further refine the regression of boundary distances, the authors developed two specialized modules:
- Surroundings Cross-Correlation Sensitive (SCCS): This module enables the model to analyze the neighborhood of a pixel, providing better context for regressing the distance to the object boundary.
- Residual IoU (R-IoU) Loss: Unlike previous models that optimized the ratio of predicted to actual distances, R-IoU focuses on the raw distance residual. This mathematical simplification leads to more stable training and faster convergence.
Experimental Results
Performance on Standard Benchmarks
VeinMask was evaluated on two primary benchmarks: the Common Objects in Context (COCO) test-dev and the Semantic Boundaries Dataset (SBD). On the COCO dataset, VeinMask achieved a 32.4% mean Average Precision (mAP). Notably, it outperformed existing one-stage contour methods while utilizing only half the design complexity (8 directions versus 16 or more in competing models).
Ablation Studies
The effectiveness of the hierarchical growth model was confirmed through ablation studies on the SBD dataset. The inclusion of the minor vein refinement process alone yielded a 3.1% mAP improvement, proving that the bio-inspired growth mechanism is highly effective for capturing complex boundaries.
Module Portability
A significant finding of this research is the "plug-and-play" nature of the Centroidness and R-IoU modules. When integrated into the older PolarMask framework, these modules boosted performance by 2.4% mAP without increasing computational cost, suggesting broad applicability across other contour-based architectures.
Discussion and Conclusion
VeinMask demonstrates that dynamic complexity is a powerful tool for shape representation. By mimicking the hierarchical structure of a leaf, the model achieves a 92% mask cover ratio with minimal directional complexity. This efficiency stems from the realization that not all parts of a contour require the same level of detail.
Despite its successes, the model faces limitations regarding small objects. Because VeinMask relies on predicting vein lengths, minor errors in short veins can result in significant proportional distortions in the final mask. Future research will focus on enhancing the robustness of vein growth for tiny instances and exploring the integration of this logic into multi-stage refinement pipelines.
In summary, VeinMask provides a novel, efficient, and biologically inspired framework that sets a new standard for one-stage instance segmentation.
Summary updated