Growing Instance Mask on Leaf

TL;DR

VeinMask is a novel single-shot instance segmentation method inspired by leaf vein growth, achieving high accuracy with low complexity by predicting veins in polar coordinates and introducing new modules and losses.

Contribution

The paper introduces VeinMask, a simple yet effective contour-based segmentation approach that models vein growth, predicts in polar coordinates, and incorporates new modules for improved performance.

Findings

Outperforms existing contour-based methods on COCO dataset.

Achieves high accuracy with nearly half the design complexity.

Effective supervision with Residual IoU loss enhances regression tasks.

Abstract

Contour-based instance segmentation methods include one-stage and multi-stage schemes. These approaches achieve remarkable performance. However, they have to define plenty of points to segment precise masks, which leads to high complexity. We follow this issue and present a single-shot method, called \textbf{VeinMask}, for achieving competitive performance in low design complexity. Concretely, we observe that the leaf locates coarse margins via major veins and grows minor veins to refine twisty parts, which makes it possible to cover any objects accurately. Meanwhile, major and minor veins share the same growth mode, which avoids modeling them separately and ensures model simplicity. Considering the superiorities above, we propose VeinMask to formulate the instance segmentation problem as the simulation of the vein growth process and to predict the major and minor veins in polar coordinates. Besides, centroidness is introduced for instance segmentation tasks to help suppress low-quality instances. Furthermore, a surroundings cross-correlation sensitive (SCCS) module is designed to enhance the feature expression by utilizing the surroundings of each pixel. Additionally, a Residual IoU (R-IoU) loss is formulated to supervise the regression tasks of major and minor veins effectively. Experiments demonstrate that VeinMask performs much better than other contour-based methods in low design complexity. Particularly, our method outperforms existing one-stage contour-based methods on the COCO dataset with almost half the design complexity.