Box Gaussian Loss in Rotated Object Detection

• Box Gaussian Loss is a framework that represents boxes as Gaussians, enabling optimization via discrepancies like Wasserstein, KL, and Bhattacharyya for robust rotated detection.
• The method alleviates direct regression challenges such as angle periodicity and edge-swapping by encoding box geometry through mean and covariance mappings.
• Empirical studies demonstrate that BGL enhances average precision in both 2D and 3D detection tasks and extends to applications like taxonomy expansion.

Box Gaussian Loss (BGL) denotes a family of objectives that map a box representation to a Gaussian distribution and then optimize a discrepancy between the resulting Gaussians rather than between raw box parameters. In computer vision, this design emerged chiefly for rotated object detection, where direct regression on $(x,y,w,h,\theta)$ is poorly aligned with rotated IoU and is vulnerable to angle periodicity, edge-swapping ambiguity, and zero-overlap failure modes. Across the literature, BGL has been instantiated with Gaussian Wasserstein distance, Bhattacharyya- and Hellinger-based Probabilistic Intersection-over-Union, Kullback–Leibler divergence, and related Gaussian metrics; in some papers the term “Box Gaussian Loss” is explicit, whereas in others the same construction is described as a “Gaussian-based loss” without the BGL label (Yang et al., 2021, Llerena et al., 2021, Yang et al., 2022, Xiong et al., 19 Sep 2025, Mishra et al., 14 Jan 2026).

1. Terminological scope and historical usage

The expression “Box Gaussian Loss” is not attached to a single canonical formula. Rather, it refers to the general procedure of representing a box as a Gaussian and regressing in Gaussian space. The 2021 paper on rotated object detection via Gaussian Wasserstein distance states that its core idea is exactly a Box Gaussian Loss, even though the paper does not explicitly use the name “BGL” (Yang et al., 2021). By contrast, the 2021 paper on Gaussian bounding boxes and ProbIoU explicitly defines two concrete BGL variants—BGL-H and BGL-B—derived from Hellinger distance and Bhattacharyya distance, respectively (Llerena et al., 2021). The 2022 paper on detecting rotated objects as Gaussian distributions likewise does not use the BGL name, but treats KLD-, BCD-, and GWD-based Gaussian losses as interchangeable members of the same methodological family (Yang et al., 2022).

Paper	Gaussian discrepancy	Terminology in paper
"Rethinking Rotated Object Detection with Gaussian Wasserstein Distance Loss" (Yang et al., 2021)	Gaussian Wasserstein distance	Not explicitly named BGL
"Gaussian Bounding Boxes and Probabilistic Intersection-over-Union for Object Detection" (Llerena et al., 2021)	Hellinger / Bhattacharyya / ProbIoU	BGL-H and BGL-B
"Detecting Rotated Objects as Gaussian Distributions and Its 3-D Generalization" (Yang et al., 2022)	KLD, BCD, GWD	“Gaussian-based losses”
"RadarGaussianDet3D: An Efficient and Effective Gaussian-based 3D Detector with 4D Automotive Radars" (Xiong et al., 19 Sep 2025)	3D Gaussian KL divergence	Explicitly named BGL
"Enhancing Rotated Object Detection via Anisotropic Gaussian Bounding Box and Bhattacharyya Distance" (Thai et al., 18 Oct 2025)	Bhattacharyya distance	Gaussian box loss, not explicitly BGL
"TaxoBell: Gaussian Box Embeddings for Self-Supervised Taxonomy Expansion" (Mishra et al., 14 Jan 2026)	BC-, KL-, and regularization-based energies	Not explicitly named BGL

This terminological diversity is substantive rather than cosmetic. Different papers adopt different covariance parameterizations, different Gaussian discrepancies, different normalization schemes, and different interpretations of what properties matter most: symmetry, scale invariance, metric structure, or high-IoU alignment. A precise use of “BGL” therefore requires specifying both the box-to-Gaussian mapping and the Gaussian discrepancy being optimized.

2. Box-to-Gaussian parameterization

The common core of BGL is the conversion of a box into a Gaussian with mean given by the box center and covariance encoding size and orientation. For a 2D oriented bounding box $b=(x,y,w,h,\theta)$, one family of methods uses

$$\mu = [x,y]^\top,\qquad R(\theta)= \begin{bmatrix} \cos\theta & -\sin\theta\ \sin\theta & \cos\theta \end{bmatrix},$$

with covariance derived from the second moments of a uniform rectangle:

$$\Sigma = R(\theta)\operatorname{diag}(w^2/12,\;h^2/12)R(\theta)^\top.$$

This is the construction used in the ProbIoU framework, where the Gaussian is treated as a “fuzzy representation” of the object region and can be rendered as an ellipse with area-equivalence constant $r=\sqrt{12/\pi}$ (Llerena et al., 2021).

A second lineage, used by the GWD and KLD-based rotated detection papers, defines a symmetric “square-root covariance”

$$\Sigma^{1/2}=R(\theta)\operatorname{diag}(w/2,\;h/2)R(\theta)^\top,$$

and then

$$\Sigma=\Sigma^{1/2}\Sigma^{1/2\top} =R(\theta)\operatorname{diag}((w/2)^2,\;(h/2)^2)R(\theta)^\top.$$

This choice is not the conventional second-moment covariance of a uniform rectangle; the 2021 GWD paper explicitly notes that the paper uses $(w/2)^2$ and $(h/2)^2$ rather than $w^2/12$ and $b=(x,y,w,h,\theta)$0, because the commutative horizontal case then reduces to a familiar $b=(x,y,w,h,\theta)$1-norm regularization on widths and heights (Yang et al., 2021). The 2022 KLD paper adopts the same half-width/half-height convention and states three properties of the resulting $b=(x,y,w,h,\theta)$2: edge exchangeability,

$b=(x,y,w,h,\theta)$3

angle periodicity,

$b=(x,y,w,h,\theta)$4

and approximate angle invariance when $b=(x,y,w,h,\theta)$5 (Yang et al., 2022).

The same principle extends to 3D. RadarGaussianDet3D maps a 3D box $b=(x,y,w,h,\theta)$6 to a Gaussian with

$b=(x,y,w,h,\theta)$7

where

$b=(x,y,w,h,\theta)$8

Here $b=(x,y,w,h,\theta)$9 is a class-dependent scaling hyperparameter, set to $$\mu = [x,y]^\top,\qquad R(\theta)= \begin{bmatrix} \cos\theta & -\sin\theta\ \sin\theta & \cos\theta \end{bmatrix},$$0 for pedestrians and cyclists and $$\mu = [x,y]^\top,\qquad R(\theta)= \begin{bmatrix} \cos\theta & -\sin\theta\ \sin\theta & \cos\theta \end{bmatrix},$$1 for cars and trucks in the reported experiments (Xiong et al., 19 Sep 2025).

3. Principal loss formulations

Gaussian Wasserstein distance. In the GWD formulation, the discrepancy between $$\mu = [x,y]^\top,\qquad R(\theta)= \begin{bmatrix} \cos\theta & -\sin\theta\ \sin\theta & \cos\theta \end{bmatrix},$$2 and $$\mu = [x,y]^\top,\qquad R(\theta)= \begin{bmatrix} \cos\theta & -\sin\theta\ \sin\theta & \cos\theta \end{bmatrix},$$3 is the squared 2-Wasserstein distance

$$\mu = [x,y]^\top,\qquad R(\theta)= \begin{bmatrix} \cos\theta & -\sin\theta\ \sin\theta & \cos\theta \end{bmatrix},$$4

The covariance term is the Bures metric. Because raw $$\mu = [x,y]^\top,\qquad R(\theta)= \begin{bmatrix} \cos\theta & -\sin\theta\ \sin\theta & \cos\theta \end{bmatrix},$$5 grows quickly for large errors, the paper uses an IoU-like transformed loss

$$\mu = [x,y]^\top,\qquad R(\theta)= \begin{bmatrix} \cos\theta & -\sin\theta\ \sin\theta & \cos\theta \end{bmatrix},$$6

with $$\mu = [x,y]^\top,\qquad R(\theta)= \begin{bmatrix} \cos\theta & -\sin\theta\ \sin\theta & \cos\theta \end{bmatrix},$$7 or $$\mu = [x,y]^\top,\qquad R(\theta)= \begin{bmatrix} \cos\theta & -\sin\theta\ \sin\theta & \cos\theta \end{bmatrix},$$8. The paper reports that direct use of $$\mu = [x,y]^\top,\qquad R(\theta)= \begin{bmatrix} \cos\theta & -\sin\theta\ \sin\theta & \cos\theta \end{bmatrix},$$9 is too sensitive to outliers, while the appendix variant using $$\Sigma = R(\theta)\operatorname{diag}(w^2/12,\;h^2/12)R(\theta)^\top.$$0 alone with $$\Sigma = R(\theta)\operatorname{diag}(w^2/12,\;h^2/12)R(\theta)^\top.$$1 slightly improves accuracy and reduces hyperparameters (Yang et al., 2021).

ProbIoU and the explicit BGL nomenclature. The ProbIoU paper defines the Bhattacharyya coefficient

$$\Sigma = R(\theta)\operatorname{diag}(w^2/12,\;h^2/12)R(\theta)^\top.$$2

the Hellinger distance

$$\Sigma = R(\theta)\operatorname{diag}(w^2/12,\;h^2/12)R(\theta)^\top.$$3

and

$$\Sigma = R(\theta)\operatorname{diag}(w^2/12,\;h^2/12)R(\theta)^\top.$$4

Two localization losses are then defined. BGL-H is

$$\Sigma = R(\theta)\operatorname{diag}(w^2/12,\;h^2/12)R(\theta)^\top.$$5

and BGL-B is

$$\Sigma = R(\theta)\operatorname{diag}(w^2/12,\;h^2/12)R(\theta)^\top.$$6

The experiments use a two-stage schedule: train with $$\Sigma = R(\theta)\operatorname{diag}(w^2/12,\;h^2/12)R(\theta)^\top.$$7 for half the iterations and then switch to $$\Sigma = R(\theta)\operatorname{diag}(w^2/12,\;h^2/12)R(\theta)^\top.$$8, combining localization and classification as

$$\Sigma = R(\theta)\operatorname{diag}(w^2/12,\;h^2/12)R(\theta)^\top.$$9

with $r=\sqrt{12/\pi}$0 often used to keep gradient magnitudes consistent when switching (Llerena et al., 2021).

KLD-, BCD-, and normalized Gaussian losses. The 2022 rotated detection paper treats BGL as a Gaussian regression loss instantiated by classical distances between multivariate Gaussians, most effectively KLD. For predicted and target Gaussians $r=\sqrt{12/\pi}$1 and $r=\sqrt{12/\pi}$2, it uses the closed-form

$r=\sqrt{12/\pi}$3

The same paper also analyzes GWD and Bhattacharyya distance and states that, for KLD and BCD, the same IoU-like normalization pattern used for GWD can be applied, integrating the regression term with focal loss via

$r=\sqrt{12/\pi}$4

with default $r=\sqrt{12/\pi}$5 (Yang et al., 2022).

Bhattacharyya-distance refinements. A later rotated detection paper formulates a Bhattacharyya Gaussian Box Loss using

$r=\sqrt{12/\pi}$6

with $r=\sqrt{12/\pi}$7, followed by the calibrated score

$r=\sqrt{12/\pi}$8

The same work introduces an anisotropic Gaussian mapping for near-square boxes by modulating eigenvalues with $r=\sqrt{12/\pi}$9 and sets $$\Sigma^{1/2}=R(\theta)\operatorname{diag}(w/2,\;h/2)R(\theta)^\top,$$0 in the reported experiments (Thai et al., 18 Oct 2025).

3D KL-based BGL. RadarGaussianDet3D explicitly names its loss BGL and defines it as the average KL divergence between predicted and ground-truth 3D Gaussians:

$$\Sigma^{1/2}=R(\theta)\operatorname{diag}(w/2,\;h/2)R(\theta)^\top,$$1

where

$$\Sigma^{1/2}=R(\theta)\operatorname{diag}(w/2,\;h/2)R(\theta)^\top,$$2

This term augments the detector’s original regression loss:

$$\Sigma^{1/2}=R(\theta)\operatorname{diag}(w/2,\;h/2)R(\theta)^\top,$$3

with $$\Sigma^{1/2}=R(\theta)\operatorname{diag}(w/2,\;h/2)R(\theta)^\top,$$4 (Xiong et al., 19 Sep 2025).

Beyond detection. In TaxoBell, Gaussian box embeddings are trained with a compound energy consisting of a symmetric overlap term based on the Bhattacharyya coefficient, an asymmetric containment term based on forward KL, a reverse-KL coverage constraint, and covariance regularizers. The paper does not use the phrase “Box Gaussian Loss,” but this objective is explicitly characterized as precisely what would be termed a BGL in that setting (Mishra et al., 14 Jan 2026).

4. Optimization geometry and the rotated-box problem

The original motivation for Gaussian box losses in rotated detection is the failure of conventional box-parameter regression to respect rotated-box geometry. The GWD paper identifies four bottlenecks: boundary discontinuity induced by periodic angle parameterizations, inconsistency between smooth $$\Sigma^{1/2}=R(\theta)\operatorname{diag}(w/2,\;h/2)R(\theta)^\top,$$5 regression and IoU-based evaluation, the square-like ambiguity when $$\Sigma^{1/2}=R(\theta)\operatorname{diag}(w/2,\;h/2)R(\theta)^\top,$$6, and the absence of informative gradients when rotated boxes do not overlap (Yang et al., 2021). In OpenCV-style and long-edge parameterizations, the same physical rectangle can be represented by angle shifts and width-height swaps, so direct angle regression creates loss jumps at representational boundaries.

The Gaussian representation removes much of this pathology because the loss depends on $$\Sigma^{1/2}=R(\theta)\operatorname{diag}(w/2,\;h/2)R(\theta)^\top,$$7 rather than on $$\Sigma^{1/2}=R(\theta)\operatorname{diag}(w/2,\;h/2)R(\theta)^\top,$$8 directly. Under the half-extent mapping, $$\Sigma^{1/2}=R(\theta)\operatorname{diag}(w/2,\;h/2)R(\theta)^\top,$$9 and $$\Sigma=\Sigma^{1/2}\Sigma^{1/2\top} =R(\theta)\operatorname{diag}((w/2)^2,\;(h/2)^2)R(\theta)^\top.$$0 yield identical $$\Sigma=\Sigma^{1/2}\Sigma^{1/2\top} =R(\theta)\operatorname{diag}((w/2)^2,\;(h/2)^2)R(\theta)^\top.$$1, while $$\Sigma=\Sigma^{1/2}\Sigma^{1/2\top} =R(\theta)\operatorname{diag}((w/2)^2,\;(h/2)^2)R(\theta)^\top.$$2 and $$\Sigma=\Sigma^{1/2}\Sigma^{1/2\top} =R(\theta)\operatorname{diag}((w/2)^2,\;(h/2)^2)R(\theta)^\top.$$3 yield identical $$\Sigma=\Sigma^{1/2}\Sigma^{1/2\top} =R(\theta)\operatorname{diag}((w/2)^2,\;(h/2)^2)R(\theta)^\top.$$4 as well. This makes the loss invariant to the parameter wrap-around that produces boundary discontinuities in raw angle regression, and it explains why square-like boxes, whose covariance is nearly isotropic, no longer incur large angle penalties merely for representational reasons (Yang et al., 2022).

The gradient structure is one of the strongest arguments for KLD-based BGL. The 2022 paper shows that center gradients are scaled by inverse target sizes, so small targets induce larger gradients and offsets along the short edge are penalized more than offsets along the long edge. The angle gradient

$$\Sigma=\Sigma^{1/2}\Sigma^{1/2\top} =R(\theta)\operatorname{diag}((w/2)^2,\;(h/2)^2)R(\theta)^\top.$$5

becomes larger as aspect ratio increases, which the paper interprets as a self-modulated update rule with clear physical meaning for high-IoU localization (Yang et al., 2022).

Not all Gaussian losses behave identically in this geometry. The ProbIoU paper emphasizes that Gaussian overlap remains strictly positive because Gaussians have infinite support, and that BGL-B avoids vanishing gradients at large separations precisely because it is unbounded above (Llerena et al., 2021). By contrast, the later Bhattacharyya-distance paper argues that the isotropic behavior of square-like covariances can under-penalize angle errors and therefore introduces anisotropy via $$\Sigma=\Sigma^{1/2}\Sigma^{1/2\top} =R(\theta)\operatorname{diag}((w/2)^2,\;(h/2)^2)R(\theta)^\top.$$6 to make intermediate orientations distinguishable while preserving $$\Sigma=\Sigma^{1/2}\Sigma^{1/2\top} =R(\theta)\operatorname{diag}((w/2)^2,\;(h/2)^2)R(\theta)^\top.$$7 periodic structure (Thai et al., 18 Oct 2025). This development suggests that “rotation invariance” in BGL is not an unconditional virtue; in some regimes it must be selectively relaxed.

5. Empirical performance in 2D and 3D detection

The earliest large-scale evidence for BGL-style regression in rotated detection comes from the GWD paper, which evaluates five datasets and multiple detectors. On DOTA, RetinaNet with $$\Sigma=\Sigma^{1/2}\Sigma^{1/2\top} =R(\theta)\operatorname{diag}((w/2)^2,\;(h/2)^2)R(\theta)^\top.$$8 improves from $$\Sigma=\Sigma^{1/2}\Sigma^{1/2\top} =R(\theta)\operatorname{diag}((w/2)^2,\;(h/2)^2)R(\theta)^\top.$$9 to $(w/2)^2$0 mAP50 and RetinaNet with $(w/2)^2$1 improves from $(w/2)^2$2 to $(w/2)^2$3 mAP50; R$(w/2)^2$4Det with $(w/2)^2$5 improves from $(w/2)^2$6 to $(w/2)^2$7 mAP50. On HRSC2016 ships, high-IoU behavior improves strongly: RetinaNet AP75 rises from $(w/2)^2$8 to $(w/2)^2$9, and R$(h/2)^2$0Det AP75 rises from $(h/2)^2$1 to $(h/2)^2$2. Additional gains are reported on UCAS-AOD, ICDAR2015, and ICDAR2017 MLT, and with training strategies and tricks the method reaches approximately $(h/2)^2$3 mAP50 overall on DOTA (Yang et al., 2021).

The ProbIoU paper extends this evidence beyond rotated aerial benchmarks. On PASCAL VOC 2007, EfficientDet D0 with the two-stage BGL achieves AP (IoU metric) approximately $(h/2)^2$4 and AP (ProbIoU metric) approximately $(h/2)^2$5, while SSD300 reaches AP (ProbIoU) approximately $(h/2)^2$6 and AP75 (ProbIoU) approximately $(h/2)^2$7. On rotated detection, DOTA v1 with RetinaNet R-50 yields average AP50 approximately $(h/2)^2$8 for BGL versus $(h/2)^2$9 for re-trained GWD, and on HRSC2016 the reported AP50:95 is approximately $w^2/12$0 for R$w^2/12$1Det under the reduced-batch setting discussed in the paper (Llerena et al., 2021).

The KLD-centered Gaussian-loss study broadens the evaluation to twelve public datasets across 2D and 3D. On DOTA-v1.0, RetinaNet plus KLD or BCD raises AP50 from $w^2/12$2 under Smooth L1 to approximately $w^2/12$3 and $w^2/12$4, respectively. On HRSC2016, the reported AP75 gains are especially large: RetinaNet + KLD improves by $w^2/12$5, and R$w^2/12$6Det + KLD improves by $w^2/12$7. On DIOR-R, KLD reaches $w^2/12$8 mAP50, and with multiscale augmentation and training/testing the paper reports $w^2/12$9 for RetinaNet and $b=(x,y,w,h,\theta)$00 for FPN (Yang et al., 2022).

The Bhattacharyya-distance and anisotropic Gaussian paper reports further gains on DOTA v1.0. For RetinaNet, $b=(x,y,w,h,\theta)$01 achieves $b=(x,y,w,h,\theta)$02 AP50 versus $b=(x,y,w,h,\theta)$03 for SmoothL1, $b=(x,y,w,h,\theta)$04 for GWD, $b=(x,y,w,h,\theta)$05 for KLD, and $b=(x,y,w,h,\theta)$06 for KFIoU. For R3Det, $b=(x,y,w,h,\theta)$07 reaches $b=(x,y,w,h,\theta)$08 AP50 versus $b=(x,y,w,h,\theta)$09 for SmoothL1, $b=(x,y,w,h,\theta)$10 for GWD, $b=(x,y,w,h,\theta)$11 for KLD, and $b=(x,y,w,h,\theta)$12 for KFIoU. The anisotropic Gaussian Bounding Box (AGBB) variant is reported to improve AP75 and mAP, with R3Det mAP rising from $b=(x,y,w,h,\theta)$13 under GBB to $b=(x,y,w,h,\theta)$14 under AGBB (Thai et al., 18 Oct 2025).

In 3D radar detection, RadarGaussianDet3D isolates the contribution of its explicitly named BGL in ablation. Without BGL, the full PGE system achieves $b=(x,y,w,h,\theta)$15 mAP (3D) and $b=(x,y,w,h,\theta)$16 mAP (BEV) on TJ4DRadSet; adding BGL increases these to $b=(x,y,w,h,\theta)$17 and $b=(x,y,w,h,\theta)$18, corresponding to gains of $b=(x,y,w,h,\theta)$19 in 3D mAP and $b=(x,y,w,h,\theta)$20 in BEV mAP. The paper further states that the full system achieves state-of-the-art detection accuracy while delivering substantially faster inference, although that full-system claim includes both PGE and BGL rather than BGL alone (Xiong et al., 19 Sep 2025).

6. Extensions, limitations, and recurring misconceptions

BGL is not restricted to visual object detection. TaxoBell applies Gaussian box embeddings to taxonomy expansion, representing concepts by axis-aligned boxes mapped to diagonal Gaussians and training with a combined symmetric-overlap and asymmetric-containment objective. On five benchmark datasets, the paper reports that TaxoBell outperforms eight state-of-the-art taxonomy expansion baselines by $b=(x,y,w,h,\theta)$21 in MRR and around $b=(x,y,w,h,\theta)$22 in Recall@k, with MR reductions up to around $b=(x,y,w,h,\theta)$23 (Mishra et al., 14 Jan 2026). This demonstrates that the BGL idea generalizes from localization geometry to hierarchical reasoning whenever boxes and Gaussian uncertainty can be jointly interpreted.

A persistent misconception is that BGL denotes a single loss with a fixed set of invariances. The literature shows otherwise. GWD is described as symmetric, positive, zero iff equal, and differentiable, with a clear decomposition into mean Euclidean distance plus Bures covariance distance (Yang et al., 2021). ProbIoU is symmetric and scale-invariant, and BGL-B is favored when non-vanishing gradients at large separations are needed (Llerena et al., 2021). The 2022 Gaussian-loss paper explicitly states that KLD and BCD are scale-invariant, while GWD is not scale-invariant and is sensitive to large center offsets unless normalized (Yang et al., 2022). Consequently, symmetry, metric structure, scale invariance, and gradient behavior depend on the chosen Gaussian discrepancy rather than on the generic BGL label.

The box-to-Gaussian mapping is likewise non-canonical. One line of work uses the uniform-rectangle second moments $b=(x,y,w,h,\theta)$24 and $b=(x,y,w,h,\theta)$25; another uses the squared half-extents $b=(x,y,w,h,\theta)$26 and $b=(x,y,w,h,\theta)$27. The former is motivated by geometric moments, while the latter is motivated by optimization behavior and by the reduction to a familiar horizontal $b=(x,y,w,h,\theta)$28 form in the commutative case. This suggests that BGL is best understood as a modeling framework with multiple legitimate parameterizations rather than a single mathematically forced construction (Llerena et al., 2021, Yang et al., 2022).

The main limitations reported so far are technical rather than conceptual. The GWD paper notes that very small $b=(x,y,w,h,\theta)$29 can create ill-conditioned covariances and recommends clamping widths and heights and optionally adding $b=(x,y,w,h,\theta)$30 before square roots; it also remarks that metric alignment to rotated IoU is approximate rather than exact (Yang et al., 2021). The 3D radar paper uses a fixed class-dependent scaling factor $b=(x,y,w,h,\theta)$31, and explicitly notes that learning uncertainty or directly learning covariance scales could tailor Mahalanobis weighting more adaptively (Xiong et al., 19 Sep 2025). The anisotropic Bhattacharyya paper states that a single Gaussian cannot represent multimodal uncertainty in cluttered or occluded scenes and proposes mixture-of-Gaussians as a possible extension (Thai et al., 18 Oct 2025). The 2022 Gaussian-loss paper adds that the approach does not directly extend to polygonal or quadrilateral boxes, and that square-like BEV boxes in 3D induce heading ambiguity that required a dedicated post-processing step in its 3D generalization (Yang et al., 2022).

Taken together, these results define BGL less as a single loss function than as a distributional viewpoint on boxes. A box becomes a Gaussian whose mean encodes location and whose covariance encodes extent, rotation, and, in some applications, uncertainty or semantic generality. The choice of Gaussian discrepancy—Wasserstein, Bhattacharyya, Hellinger, KL, or a compound energy—then determines the exact inductive bias: metric symmetry, scale sensitivity, overlap behavior, high-IoU alignment, and numerical robustness. In that broader sense, BGL is a unifying language for replacing parameter-space regression with geometry-aware optimization in Gaussian box space.