Normalized Gaussian Wasserstein Distance

• Normalized Gaussian Wasserstein Distance (NWD) is a similarity measure that converts the 2-Wasserstein distance between Gaussian-modeled bounding boxes into a normalized score in (0,1] to improve tiny object detection.
• It utilizes a box-to-Gaussian mapping where an axis-aligned bounding box is represented by its center and scaled covariance, ensuring smooth sensitivity to positional and scale variations.
• Empirical studies show that integrating NWD in detection pipelines significantly boosts average precision by providing robust supervision where traditional IoU metrics fail.

Searching arXiv for the cited NWD and related Gaussian Wasserstein papers to ground the article in current arXiv records. Normalized Gaussian Wasserstein Distance (NWD) is a Gaussianized similarity measure derived from the $2$-Wasserstein distance between probability distributions and used most prominently for tiny object detection. In its canonical object-detection form, an axis-aligned bounding box $R=(cx,cy,w,h)$ is modeled as a $2$D Gaussian whose mean is the box center and whose covariance encodes box size; the Gaussian $2$-Wasserstein distance between two such boxes is then mapped through an exponential normalization into a score in $(0,1]$, with larger values indicating greater similarity (Wang et al., 2021). Related work uses “normalized Wasserstein” in other senses—most notably normalization over mixture proportions rather than spatial scale—so the term is context-dependent and should not be treated as a single universal construction (Balaji et al., 2019).

1. Definition and canonical formulation

In the tiny-object detection literature, NWD begins with a box-to-Gaussian mapping. For an axis-aligned bounding box $R=(cx, cy, w, h)$, the Gaussian parameters are

$$\boldsymbol{\mu} = \begin{bmatrix} cx \ cy \end{bmatrix}, \quad \boldsymbol{\Sigma} = \begin{bmatrix} \frac{w^2}{4} & 0 \ 0 & \frac{h^2}{4} \end{bmatrix}.$$

This representation follows from identifying the box with an inscribed ellipse whose semi-axes are $w/2$ and $h/2$, then matching that ellipse to a Gaussian level set (Wang et al., 2021).

The underlying geometric quantity is the squared Gaussian $2$-Wasserstein distance. For Gaussians $R=(cx,cy,w,h)$0 and $R=(cx,cy,w,h)$1, the general closed form is

$R=(cx,cy,w,h)$2

and for the diagonal covariances used for horizontal boxes this simplifies to

$R=(cx,cy,w,h)$3

Hence the Gaussian Wasserstein distance between two axis-aligned boxes reduces to an Euclidean distance in a $R=(cx,cy,w,h)$4-dimensional parameter space consisting of center coordinates and half-sizes (Wang et al., 2021).

The normalized similarity is then defined as

$R=(cx,cy,w,h)$5

where $R=(cx,cy,w,h)$6 is a dataset-dependent constant. In the AI-TOD setting, $R=(cx,cy,w,h)$7 is empirically set to the average absolute object size in the dataset; for AI-TOD-v2, the corresponding paper reports a mean size of approximately $R=(cx,cy,w,h)$8 pixels and uses that scale as a practical default (Wang et al., 2021, Xu et al., 2022). Because $R=(cx,cy,w,h)$9, NWD lies in $2$0, equals $2$1 for identical boxes, and decays toward $2$2 as the boxes become more dissimilar (Wang et al., 2021).

2. Mathematical basis in Gaussian optimal transport

NWD inherits its geometry from Gaussian optimal transport rather than from set overlap. The relevant base object is the $2$3-Wasserstein distance, whose Gaussian closed form is standard and appears both in Gaussian word embeddings and in statistical analyses of Gaussian Wasserstein functionals (Sun et al., 2018, Rippl et al., 2015). In the Gaussian case, the distance depends jointly on mean displacement and covariance mismatch: $2$4 For the diagonal box covariances used in NWD, the covariance term collapses to squared differences of half-widths and half-heights, yielding the $2$5-D Euclidean form above (Wang et al., 2021).

This Gaussian OT basis is central to why NWD behaves differently from IoU. Wasserstein distance remains meaningful when supports do not overlap, and it changes continuously with center and scale perturbations. The tiny-object detection papers explicitly emphasize that this produces informative similarity values even when two boxes are disjoint or one fully contains the other, regimes in which IoU is either exactly zero or poorly discriminative (Wang et al., 2021, Xu et al., 2022).

The literature also makes clear that “normalization” is not intrinsic to the Gaussian $2$6 formula itself. In Gaussian word embedding, the model uses the raw $2$7 between Gaussian word representations and only applies an external affine transformation $2$8 so that the distance can enter a logistic loss; the paper does not define a separately named NWD (Sun et al., 2018). A plausible implication is that the normalization in NWD is task-specific rather than a universal mathematical property of Gaussian Wasserstein distance.

3. Motivation from tiny object detection

The primary motivation for NWD is the failure mode of IoU-based similarity metrics on tiny objects. The object-detection papers analyze this in terms of discrete pixel geometry: for a tiny box, a shift of only a few pixels can cause IoU to collapse, whereas the same absolute shift produces only a modest IoU change for a much larger box (Wang et al., 2021). One paper states that for a tiny $2$9 box, moving the prediction by a small amount changes IoU from $2$0 to $2$1, while for a normal $2$2 box the same pixel shift changes IoU from $2$3 to $2$4 (Wang et al., 2021). The AI-TOD dataset is described as having mean object size $2$5 pixels, with “tiny” defined as $2$6 pixels, which makes these effects operationally severe (Wang et al., 2021).

The resulting problems propagate through anchor-based detectors in three places: label assignment, non-maximum suppression, and regression loss. Under IoU-threshold assignment, tiny objects receive very few positive anchors; the AI-TOD study reports that the average number of positive anchors per ground truth is less than $2$7 under IoU (Wang et al., 2021). IoU-based NMS can fail to suppress redundant detections because slightly mislocalized tiny boxes may have IoU below the NMS threshold, and IoU-based regression losses can provide zero or poor gradients when boxes do not overlap or one box contains the other (Wang et al., 2021).

NWD is introduced as a remedy because it replaces overlap-based similarity by a smooth geometric similarity in Gaussian parameter space. The papers emphasize three properties: smoothness with respect to location deviation, scale balance across object sizes, and meaningful nonzero similarity for non-overlapping boxes (Wang et al., 2021, Xu et al., 2022). In the deviation plots discussed in both papers, NWD-deviation curves for different box scales “completely coincide,” whereas IoU-deviation curves deteriorate sharply as object size decreases (Wang et al., 2021, Xu et al., 2022). This suggests that NWD is designed to equalize supervision quality across scales rather than to approximate overlap more faithfully.

4. Detector integration and training workflows

The original tiny-object detection paper states that NWD can replace IoU in anchor assignment, NMS, and regression loss for anchor-based detectors such as Faster R-CNN, RetinaNet, ATSS, Cascade R-CNN, and DetectoRS (Wang et al., 2021). In threshold-based assignment, the same positive and negative thresholds $2$8 are retained, but IoU is replaced by NWD. For Faster R-CNN-style RPN training, an anchor is positive if it is the anchor with highest NWD to a ground truth and that NWD is at least $2$9, or if it has NWD at least $(0,1]$0 to any ground truth; it is negative if its NWD to all ground truths is below $(0,1]$1 (Wang et al., 2021). The reported average number of positive anchors per ground truth on AI-TOD is $(0,1]$2 for IoU, $(0,1]$3 for GIoU, $(0,1]$4 for DIoU, $(0,1]$5 for CIoU, and $(0,1]$6 for NWD (Wang et al., 2021).

For regression, the same paper introduces the localization loss

$(0,1]$7

which is differentiable and remains informative when $(0,1]$8 or one box fully contains the other (Wang et al., 2021). For NMS, IoU-based suppression is replaced by an NWD-based similarity comparison so that redundant predictions around tiny objects can still be removed even if their overlap is small (Wang et al., 2021).

Subsequent work argues that label assignment is the dominant bottleneck and therefore focuses on NWD primarily as an assigning metric rather than as a loss. The “NWD-RKA” method couples NWD with a RanKing-based Assigning strategy: for each ground truth, anchors are ranked by similarity, and the top-$(0,1]$9 anchors are labeled positive (Xu et al., 2022). This eliminates brittle fixed-threshold assignment and guarantees a prescribed number of positives per ground truth. The paper states that “IoU-RKA” remains limited because many nearby anchors share identical IoU $R=(cx, cy, w, h)$0 for tiny objects, whereas NWD provides a strict ordering even for non-overlapping anchors (Xu et al., 2022). On AI-TOD-v2, the recommended practical setting is $R=(cx, cy, w, h)$1, and the paper reports robustness across $R=(cx, cy, w, h)$2 as well as across $R=(cx, cy, w, h)$3 (Xu et al., 2022).

5. Empirical behavior, advantages, and criticisms

On AI-TOD, embedding NWD into Faster R-CNN yields large gains in the original proposal. When used only in assignment, AP rises from $R=(cx, cy, w, h)$4 to $R=(cx, cy, w, h)$5; when used only in NMS, AP rises from $R=(cx, cy, w, h)$6 to $R=(cx, cy, w, h)$7; when used only in loss, AP rises from $R=(cx, cy, w, h)$8 with IoU loss to $R=(cx, cy, w, h)$9 with NWD loss (Wang et al., 2021). In an ablation that replaces IoU with NWD in exactly one Faster R-CNN module, AP increases to $$\boldsymbol{\mu} = \begin{bmatrix} cx \ cy \end{bmatrix}, \quad \boldsymbol{\Sigma} = \begin{bmatrix} \frac{w^2}{4} & 0 \ 0 & \frac{h^2}{4} \end{bmatrix}.$$0 for assigning RPN, $$\boldsymbol{\mu} = \begin{bmatrix} cx \ cy \end{bmatrix}, \quad \boldsymbol{\Sigma} = \begin{bmatrix} \frac{w^2}{4} & 0 \ 0 & \frac{h^2}{4} \end{bmatrix}.$$1 for assigning RCNN, $$\boldsymbol{\mu} = \begin{bmatrix} cx \ cy \end{bmatrix}, \quad \boldsymbol{\Sigma} = \begin{bmatrix} \frac{w^2}{4} & 0 \ 0 & \frac{h^2}{4} \end{bmatrix}.$$2 for NMS RPN, $$\boldsymbol{\mu} = \begin{bmatrix} cx \ cy \end{bmatrix}, \quad \boldsymbol{\Sigma} = \begin{bmatrix} \frac{w^2}{4} & 0 \ 0 & \frac{h^2}{4} \end{bmatrix}.$$3 for loss RPN, and $$\boldsymbol{\mu} = \begin{bmatrix} cx \ cy \end{bmatrix}, \quad \boldsymbol{\Sigma} = \begin{bmatrix} \frac{w^2}{4} & 0 \ 0 & \frac{h^2}{4} \end{bmatrix}.$$4 for loss RCNN, with the largest gain arising from RPN assignment (Wang et al., 2021). The same paper reports end-to-end improvements on AI-TOD of $$\boldsymbol{\mu} = \begin{bmatrix} cx \ cy \end{bmatrix}, \quad \boldsymbol{\Sigma} = \begin{bmatrix} \frac{w^2}{4} & 0 \ 0 & \frac{h^2}{4} \end{bmatrix}.$$5 AP for RetinaNet, $$\boldsymbol{\mu} = \begin{bmatrix} cx \ cy \end{bmatrix}, \quad \boldsymbol{\Sigma} = \begin{bmatrix} \frac{w^2}{4} & 0 \ 0 & \frac{h^2}{4} \end{bmatrix}.$$6 AP for Faster R-CNN, $$\boldsymbol{\mu} = \begin{bmatrix} cx \ cy \end{bmatrix}, \quad \boldsymbol{\Sigma} = \begin{bmatrix} \frac{w^2}{4} & 0 \ 0 & \frac{h^2}{4} \end{bmatrix}.$$7 AP for Cascade R-CNN, and $$\boldsymbol{\mu} = \begin{bmatrix} cx \ cy \end{bmatrix}, \quad \boldsymbol{\Sigma} = \begin{bmatrix} \frac{w^2}{4} & 0 \ 0 & \frac{h^2}{4} \end{bmatrix}.$$8 AP for DetectoRS when equipped with NWD (Wang et al., 2021).

The AI-TOD-v2 study reports even larger gains from using NWD specifically in assignment. On AI-TOD-v2 with Faster R-CNN, AP increases from $$\boldsymbol{\mu} = \begin{bmatrix} cx \ cy \end{bmatrix}, \quad \boldsymbol{\Sigma} = \begin{bmatrix} \frac{w^2}{4} & 0 \ 0 & \frac{h^2}{4} \end{bmatrix}.$$9 for baseline IoU-threshold assignment to $w/2$0 when replacing the metric with NWD and to $w/2$1 when combining NWD with ranking-based assignment (Xu et al., 2022). In a direct metric comparison for assignment, the same paper reports AP values of $w/2$2 for GWD, $w/2$3 for GIoU, $w/2$4 for CIoU, $w/2$5 for DIoU, $w/2$6 for IoU, and $w/2$7 for NWD (Xu et al., 2022). Cross-dataset results are also positive: for DetectoRS, the reported AP improves from $w/2$8 to $w/2$9 on AI-TOD, from $h/2$0 to $h/2$1 on VisDrone2019, and from $h/2$2 to $h/2$3 on DOTA-v2.0 when using NWD-RKA (Xu et al., 2022).

However, later work argues that NWD does not provide robust scale invariance in the strong sense needed for mixed-scale datasets. The Gaussian Combined Distance paper states that NWD incorporates a hyperparameter $h/2$4 representing the average size of objects in the dataset, but that its performance remains inconsistent across general datasets because this is a global dataset-level normalization rather than an instance-level scale normalization (Guan et al., 31 Oct 2025). The same paper further argues that NWD inherits WD’s “independent center optimization”: the center gradient structure remains essentially proportional to $h/2$5 and $h/2$6, independent of box size, so global normalization and exponential mapping do not couple center optimization with width and height (Guan et al., 31 Oct 2025). Empirically, that paper reports on AI-TOD-v2 with RetinaNet AP values of $h/2$7 for WD, $h/2$8 for NWD, and $h/2$9 for GCD; on VisDrone-2019, $2$0 for WD, $2$1 for NWD, and $2$2 for GCD; and on MS-COCO-2017, $2$3 for WD, $2$4 for NWD, $2$5 for GIoU, and $2$6 for GCD (Guan et al., 31 Oct 2025). This criticism does not negate NWD’s utility for tiny-object regimes, but it narrows the conditions under which its normalization is effective.

6. Related meanings of “normalized Wasserstein” and broader context

The term “normalized Wasserstein” is used in at least two distinct senses in the cited literature. In mixture-distribution learning, the “Normalized Wasserstein measure” is defined by optimizing over shared mixture components and separate mixture proportions for the two distributions: $2$7 Here, normalization means normalizing away mixture proportions rather than normalizing geometric scale, and the paper explicitly does not use the term “Normalized Gaussian Wasserstein Distance” or the acronym “NWD” (Balaji et al., 2019). This distinction is conceptually important because it shows that the “normalization” in NWD is application-specific.

Other cited work supplies the mathematical substrate but not the detection-specific normalization. The Gaussian word embedding paper uses raw Gaussian $2$8 as the core dissimilarity and does not introduce NWD; it instead uses a shifted energy $2$9 inside a logistic objective (Sun et al., 2018). The statistical paper on empirical Gaussian Wasserstein distance provides central limit theorems and variance-normalized statistics for Gaussian $R=(cx,cy,w,h)$00, but again does not define NWD as a geometric box similarity (Rippl et al., 2015). The paper on entropic adapted Wasserstein distance on Gaussians extends closed-form Gaussian OT to discrete-time Gaussian processes and explicitly notes that “NWD” is not used there; it discusses only possible normalized variants that one could define from the explicit formulas (Acciaio et al., 2024). These papers collectively indicate that NWD, in its established object-detection sense, is a particular normalized similarity built on Gaussian $R=(cx,cy,w,h)$01, not a universally standardized term across optimal transport.

A plausible implication is that the most precise encyclopedia usage should reserve “Normalized Gaussian Wasserstein Distance” for the box-based construction introduced for tiny object detection, while treating other “normalized Wasserstein” notions as related but distinct. Within that narrower meaning, NWD is best understood as an exponential similarity over Gaussian $R=(cx,cy,w,h)$02-Wasserstein distance with a dataset-scale constant $R=(cx,cy,w,h)$03, designed to provide smooth, scale-balanced supervision for tiny objects (Wang et al., 2021, Xu et al., 2022). Subsequent critiques show that this balance is approximate rather than absolute and motivate instance-level, jointly optimized alternatives such as Gaussian Combined Distance (Guan et al., 31 Oct 2025).