Central Weighted IoU

• Central Weighted Intersection over Union is a method that extends the standard IoU by applying a center-focused weight map to modulate pixel or region contributions.
• It adapts frameworks from semantic segmentation and bounding-box regression by replacing uniform contributions with an exponential central distance weighting scheme.
• Empirical insights indicate that central weighting can alter ranking behavior and improve robustness against noise by privileging the central region in evaluations.

Central Weighted Intersection over Union is a center-emphasized extension of Intersection over Union in which the contribution of a pixel, region, or predicted box is modulated by a weight that depends on centrality rather than being spatially uniform. In the cited literature, it appears not as a standalone benchmarked metric but as a direct adaptation of two broader frameworks: weighted IoU for semantic segmentation, where any per-pixel weight map can be inserted into the IoU computation, and IoU with smoothing for bounding box regression, where a spatially defined smoothing field augments the standard IoU loss across the whole image domain (Cho, 2021, Števuliáková et al., 2023). A central-weighted variant therefore denotes a family of IoU-based constructions in which weights are maximal near a designated center—typically an image center, object centroid, or ground-truth box center—and decay outward.

1. Foundational formulations

The standard IoU, or Jaccard Index, is defined set-theoretically as

$IoU = \frac{|A \cap A'|}{|A \cup A'|}$

for two shapes $A, A' \subseteq \mathbb{S} \subset \mathbb{R}^n$, and in semantic segmentation as

$$IoU = JI = \frac { \left| \mathbf{C}\cap \mathbf{ C }_{ gt } \right| }{ \left| \mathbf{C}\cup \mathbf{ C }_{ gt } \right| }.$$

In both cases, IoU is purely area-based: every pixel or spatial location contributes equally to the numerator and denominator (Cho, 2021).

Weighted variants retain the intersection-over-union structure while replacing uniform contributions with weighted ones. In the segmentation formulation, weighted IoU is defined as

$$wIoU=\frac { \left| \mathbf{C}\cap (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| }{ \left| \mathbf{C}\cup (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| },$$

where $\mathbf{W}(p)\in[0,1]$ is a spatial weight map and $\circ$ denotes element-wise product. The corresponding per-pixel interpretation is

$$wIoU_c = \frac{\sum_{p} W(p)\,\mathbf{1}[y_p=1 \land \hat{y}_p=1]}{\sum_{p} W(p)\,\mathbf{1}[y_p=1 \lor \hat{y}_p=1]}.$$

This construction generalizes IoU by allowing some locations to influence the score more strongly than others.

A central-weighted IoU is obtained by choosing $\mathbf{W}$ so that central locations receive larger weights than peripheral ones. This suggests that the defining feature of the central-weighted family is not a new algebraic denominator or numerator, but a particular spatial prior encoded in the weight map.

2. Segmentation-style central weighting

In the weighted-IoU framework for semantic segmentation, the original paper constructs $\mathbf{W}$ from a boundary distance map. For class $c$, foreground and background are

$A, A' \subseteq \mathbb{S} \subset \mathbb{R}^n$0

and the foreground distance-to-boundary map is

$A, A' \subseteq \mathbb{S} \subset \mathbb{R}^n$1

with Euclidean distance obtained by setting $A, A' \subseteq \mathbb{S} \subset \mathbb{R}^n$2. After normalization,

$A, A' \subseteq \mathbb{S} \subset \mathbb{R}^n$3

the paper defines the weight map as

$A, A' \subseteq \mathbb{S} \subset \mathbb{R}^n$4

where $A, A' \subseteq \mathbb{S} \subset \mathbb{R}^n$5 is the boundary importance factor (Cho, 2021).

The same framework explicitly supports a central-weighted adaptation. The cited material states that “the paper’s framework is very general: you can plug any distance-based weight map $A, A' \subseteq \mathbb{S} \subset \mathbb{R}^n$6 into Eq. \eqref{eq:wiou}.” Two central distance definitions are given. For an image or crop center $A, A' \subseteq \mathbb{S} \subset \mathbb{R}^n$7,

$A, A' \subseteq \mathbb{S} \subset \mathbb{R}^n$8

and for an object centroid,

$A, A' \subseteq \mathbb{S} \subset \mathbb{R}^n$9

These distances are then normalized to $$IoU = JI = \frac { \left| \mathbf{C}\cap \mathbf{ C }_{ gt } \right| }{ \left| \mathbf{C}\cup \mathbf{ C }_{ gt } \right| }.$$0: $$IoU = JI = \frac { \left| \mathbf{C}\cap \mathbf{ C }_{ gt } \right| }{ \left| \mathbf{C}\cup \mathbf{ C }_{ gt } \right| }.$$1

Using the same exponential template as the original weighted IoU, the central weight map is

$$IoU = JI = \frac { \left| \mathbf{C}\cap \mathbf{ C }_{ gt } \right| }{ \left| \mathbf{C}\cup \mathbf{ C }_{ gt } \right| }.$$2

and the resulting central-weighted IoU is

$$IoU = JI = \frac { \left| \mathbf{C}\cap \mathbf{ C }_{ gt } \right| }{ \left| \mathbf{C}\cup \mathbf{ C }_{ gt } \right| }.$$3

Its class-specific form is

$$IoU = JI = \frac { \left| \mathbf{C}\cap \mathbf{ C }_{ gt } \right| }{ \left| \mathbf{C}\cup \mathbf{ C }_{ gt } \right| }.$$4

This formulation preserves the exact weighted-IoU machinery while changing only the geometry of importance. Boundary wIoU weights edge-adjacent pixels most strongly; central wIoU weights the central region most strongly.

3. Bounding-box regression and spatial weighting fields

For bounding boxes, the relevant source is IoU with smoothing, which augments the standard IoU loss

$$IoU = JI = \frac { \left| \mathbf{C}\cap \mathbf{ C }_{ gt } \right| }{ \left| \mathbf{C}\cup \mathbf{ C }_{ gt } \right| }.$$5

by a smoothing penalty: $$IoU = JI = \frac { \left| \mathbf{C}\cap \mathbf{ C }_{ gt } \right| }{ \left| \mathbf{C}\cup \mathbf{ C }_{ gt } \right| }.$$6 The paper’s stated principle is “adding a smoothing part to the original IoU, where the smoothing part is given by a linear space with values that increases from the ground truth bounding box to the border of the input image, and thus covers the whole spatial search space” (Števuliáková et al., 2023).

The construction uses a normalized box representation

$$IoU = JI = \frac { \left| \mathbf{C}\cap \mathbf{ C }_{ gt } \right| }{ \left| \mathbf{C}\cup \mathbf{ C }_{ gt } \right| }.$$7

with ground-truth center $$IoU = JI = \frac { \left| \mathbf{C}\cap \mathbf{ C }_{ gt } \right| }{ \left| \mathbf{C}\cup \mathbf{ C }_{ gt } \right| }.$$8 and center-to-border normalizers

$$IoU = JI = \frac { \left| \mathbf{C}\cap \mathbf{ C }_{ gt } \right| }{ \left| \mathbf{C}\cup \mathbf{ C }_{ gt } \right| }.$$9

Right and left offsets, built from $$wIoU=\frac { \left| \mathbf{C}\cap (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| }{ \left| \mathbf{C}\cup (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| },$$0 terms, quantify how the predicted box extends beyond or lies outside the ground-truth box in each direction. These offsets define weights

$$wIoU=\frac { \left| \mathbf{C}\cap (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| }{ \left| \mathbf{C}\cup (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| },$$1

and ultimately

$$wIoU=\frac { \left| \mathbf{C}\cap (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| }{ \left| \mathbf{C}\cup (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| },$$2

so that

$$wIoU=\frac { \left| \mathbf{C}\cap (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| }{ \left| \mathbf{C}\cup (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| },$$3

The key conceptual point is that this smoothing term is a spatial weighting construction. The cited synthesis states that IoU with smoothing “creates a spatial weighting function over the image and integrates it over the predicted box area,” and that the smoothing term can indeed be interpreted as a “position-dependent weight field,” although one “that increases from the ground truth outward, not centrally peaked” (Števuliáková et al., 2023). A central-weighted IoU for boxes is therefore a natural inversion of that profile: instead of a penalty that grows toward the borders, one would use a weighting profile that is largest near the ground-truth center and decreases outward.

4. Central emphasis as a geometric design choice

The two cited frameworks imply two distinct but related notions of central weighting. In the segmentation formulation, central weighting is implemented directly as a per-pixel weight map. In the bounding-box formulation, central weighting is implemented indirectly through a spatial field or an additional penalty derived from center alignment. Both preserve the general pattern

$$wIoU=\frac { \left| \mathbf{C}\cap (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| }{ \left| \mathbf{C}\cup (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| },$$4

which the smoothing paper uses to place IoU, DIoU, CIoU, and SIoU in a common family (Števuliáková et al., 2023).

Within that family, DIoU, CIoU, and SIoU operate primarily through pairwise box relations such as center distance, aspect ratio, angle, and shape. By contrast, smoothing IoU introduces a global spatial field relative to the ground truth and image borders. The cited text summarizes the distinction as follows: DIoU, CIoU, and SIoU focus on “pairwise relations” between boxes, whereas IoU with smoothing introduces a “global spatial field” defined relative to ground truth and image borders. This is close conceptually to spatial weighting or “central weighting,” except that the implemented field is border-weighted rather than center-weighted.

A plausible implication is that “Central Weighted Intersection over Union” is best regarded as a design pattern rather than a single canonical metric. One instantiation uses the weighted-IoU equation with a central-distance map; another uses the IoU-plus-penalty template with a center-focused spatial field over the search space. In both cases, the central region becomes the dominant source of agreement or penalty.

5. Representative formulations

One representative segmentation-style construction uses the normalized central distance $$wIoU=\frac { \left| \mathbf{C}\cap (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| }{ \left| \mathbf{C}\cup (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| },$$5 together with the exponential weighting rule

$$wIoU=\frac { \left| \mathbf{C}\cap (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| }{ \left| \mathbf{C}\cup (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| },$$6

The parameter $$wIoU=\frac { \left| \mathbf{C}\cap (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| }{ \left| \mathbf{C}\cup (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| },$$7 then serves as a central importance factor. Small $$wIoU=\frac { \left| \mathbf{C}\cap (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| }{ \left| \mathbf{C}\cup (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| },$$8 yields nearly uniform weights and behavior close to standard IoU; large $$wIoU=\frac { \left| \mathbf{C}\cap (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| }{ \left| \mathbf{C}\cup (\mathbf{ C }_{ gt }\circ \mathbf{W}) \right| },$$9 concentrates the score on the central region. This mirrors the role of $\mathbf{W}(p)\in[0,1]$0 in boundary-weighted IoU, where the same exponential form continuously interpolates between nearly uniform and strongly localized weighting (Cho, 2021).

A representative bounding-box construction, proposed in the cited synthesis as an adaptation rather than as a reported result, defines a radial center-weighting function around the ground-truth center $\mathbf{W}(p)\in[0,1]$1: $\mathbf{W}(p)\in[0,1]$2 Its average over the predicted box is

$\mathbf{W}(p)\in[0,1]$3

and a corresponding loss is

$\mathbf{W}(p)\in[0,1]$4

The same source also notes a simpler “discrete version” in which central weighting is interpreted via the distance between the centers of $\mathbf{W}(p)\in[0,1]$5 and $\mathbf{W}(p)\in[0,1]$6, normalized by a maximum possible center distance (Števuliáková et al., 2023).

These formulas illustrate two different emphases. The segmentation form is a weighted evaluation metric. The box-regression form is a training loss. The common structure is that overlap is no longer judged uniformly across space; instead, central regions exert disproportionate influence.

6. Empirical context, limitations, and scope

Direct experimental validation of a central-weighted IoU is not reported in the cited papers, but the source frameworks provide relevant empirical context. For semantic segmentation, weighted IoU was validated on 33 synthetic scenes. With $\mathbf{W}(p)\in[0,1]$7 or $\mathbf{W}(p)\in[0,1]$8, wIoU was very close to standard IoU; with $\mathbf{W}(p)\in[0,1]$9 or $\circ$0, it aligned more strongly with edge-based $\circ$1, and on examples with the same total number of misclassified pixels but different spatial distributions, IoU gave exactly the same value while wIoU differentiated the segmentations (Cho, 2021). This establishes that spatial weighting can alter ranking behavior without changing the algebraic IoU template.

For bounding-box regression, IoU with smoothing was evaluated on a custom dataset of 304 images, each with exactly one object, using EfficientNetB2V2, a four-value regression head, a fixed 50% train/test split, batch size 16, 6000 iterations, and Adadelta with learning-rate decay. The compared losses were IoU, DIoU, CIoU, SIoU, and IoU plus smoothing. On clean data, IoU plus smoothing achieved train average $\circ$2, test average $\circ$3, and test best $\circ$4, outperforming the alternatives. Under 20%, 40%, and 60% synthetic box-coordinate noise, the smoothing loss remained the best-performing variant and was described as robust with respect to noise in the dimension of ground-truth bounding boxes (Števuliáková et al., 2023).

These results support two constrained inferences. First, spatially extended weighting terms can materially affect optimization and evaluation behavior. Second, the choice of weighting profile matters: boundary-centric weighting emphasizes contour fidelity, border-ramp smoothing emphasizes global guidance, and central weighting would emphasize central alignment or central-region correctness. The papers also identify limitations that carry over to central weighting. In weighted IoU, $\circ$5 must be fixed by the user; different tasks may desire different values, complicating cross-paper comparison. Shape-dependent emphasis can affect fairness. The metric was not yet explored as a loss function in the original weighted-IoU paper. For central weighting specifically, the cited synthesis notes that it should be used only where such central preference is justified, since object size, object layout, and the definition of “center” determine which regions are privileged (Cho, 2021).

Central Weighted Intersection over Union is therefore most accurately characterized as a center-prior instantiation of general weighted-IoU methodology. In segmentation, it replaces boundary distance with central distance in the weight map. In bounding-box regression, it replaces a border-increasing smoothing field with a center-peaked field or center-distance penalty. Its mathematical identity is inherited from IoU; its distinctiveness lies in how spatial importance is assigned.