Weighted Intersection over Union (wIoU) for Evaluating Image Segmentation

TL;DR

The paper tackles the inadequacy of IoU for semantic segmentation by proposing Weighted IoU ($wIoU$), a boundary-aware evaluation metric generated from a boundary-distance transform. A weight map $\mathbf{W}$, controlled by a single boundary importance factor $α$, emphasizes pixels near object boundaries while preserving interior regions, enabling a continuum between region-only and boundary-focused evaluation. The weighted IoU is defined as $wIoU=\frac{| \mathbf{C} \cap (\mathbf{C}_{gt} \circ \mathbf{W}) |}{| \mathbf{C} \cup (\mathbf{C}_{gt} \circ \mathbf{W}) |}$ and is demonstrated on a 33-scene synthetic dataset, with analyses showing small $α$ recovers IoU behavior and large $α$ aligns with edge-based metrics. The framework is argued to be robust, flexible, and applicable as both an evaluation metric and a potential training loss for semantic segmentation.

Abstract

In recent years, many semantic segmentation methods have been proposed to predict label of pixels in the scene. In general, we measure area prediction errors or boundary prediction errors for comparing methods. However, there is no intuitive evaluation metric that evaluates both aspects. In this work, we propose a new evaluation measure called weighted Intersection over Union (wIoU) for semantic segmentation. First, it builds a weight map generated from a boundary distance map, allowing weighted evaluation for each pixel based on a boundary importance factor. The proposed wIoU can evaluate both contour and region by setting a boundary importance factor. We validated the effectiveness of wIoU on a dataset of 33 scenes and demonstrated its flexibility. Using the proposed metric, we expect more flexible and intuitive evaluation in semantic segmentation field are possible.

Figures (11)

• Figure 1: Challenges in semantic segmentation and the proposed weight map. (a--b) It is difficult to predict pixel labels around object boundaries. (c) The proposed weight map emphasizes the importance of boundaries.
• Figure 2: Example of an image and its predicted classes
• Figure 3: (a) Example of a ground-truth class map $\mathbf{C}_{gt}$. (b,c) Corresponding foreground (${ \Omega }^{f}_{ c }$, white area) and background (${ \Omega }^{b}_{ c }$, shaded area) regions where $c=1$ and $c=4$, respectively. Solid blue line is the boundary of the object.
• Figure 4: Examples of distance maps. The ground-truth class map is illustrated in Figure \ref{['FIG:region']} (a).
• Figure 5: Ground truth and its weight maps according to the boundary importance factor $\alpha$. The range of weight map is $[0,1]$. White and black pixels denote weight values $1,0$, respectively. A small boundary importance factor leads a uniform weight map. On the other hand, a large boundary importance factor leads high weights around the boundaries of the objects and regions.