Shape-Aware Oriented Bounding Box (OBB) to Horizontal Bounding Box (HBB) Conversion

Badha Rathna Sabhapathy; Gotam Dahiya; Vishesh Vatsal

REVIEW 3 major objections 5 minor 10 references

Shape-Aware Oriented Bounding Box (OBB) to Horizontal Bounding Box (HBB) Conversion

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that modeling a detected ship as a superellipse and projecting that hull, rather than enclosing the oriented rectangle's corners, yields tighter horizontal bounding boxes and higher mean IoU on real remote-sensing data.

desk verdict Superellipse-based OBB-to-HBB conversion with real code and data, but the projection formula is geometrically wrong and the undershoot claims are overstated. read the letter →

arxiv 2608.05858 v1 pith:GM6DDLJU submitted 2026-08-06 cs.CV cs.RO

Badha Rathna Sabhapathy , Gotam Dahiya , Vishesh Vatsal This is my paper · ORCID

classification cs.CVcs.RO

keywords orientedboundingboxeshorizontalshipdetectionremotesensingimagerysuperellipsehullmodelboxconversionintersectionoverunionshape-awarelocalization

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that the standard way of converting an oriented bounding box (OBB) to a horizontal bounding box (HBB)—enclosing the rotated rectangle's corners—is unnecessarily loose for ships, and that a better HBB can be made by inferring a ship-shaped hull and projecting that hull onto the image axes. It models the hull as a superellipse with a shape exponent q, a fullness factor f, and a shrink factor s, derives closed-form half-extents for the axis-aligned box, and calibrates q=2.2, f=0.9, s=1.0 on synthetic hull shapes. On 10,043 real ship detections from ShipRSImageNet, the calibrated rule reaches the highest mean IoU (0.5609) among four conversion methods while reducing both overshoot and undershoot relative to the alternatives. If correct, the method makes downstream axis-aligned boxes tighter without retraining the detector, which matters for cropping, transmission, and IoU-based evaluation.

What carries the argument

The machinery is the superellipse (Lamé curve) hull model |x|^q/a^q + |y|^q/b^q = 1 in the ship's local frame, with a = L/2 and b = (W/2)·f. For a rotation θ, the proposed HBB half-extents are obtained as x_h = max_{u∈[0,1]} [a u |sin θ| + b(1-u^q)|cos θ|] and an analogous expression for y_h, where u and v interpolate between corner-like and midpoint-like contributions; because each function is concave for q>1, the optimum has a closed form. The final half-extents multiply by a shrink factor s, with explicit clipping for θ near 0° and 90°.

What would settle it

Sample the boundary of Eq. (2) densely for a ship-like a, b and θ = 30°, compute the true maximum x- and y-projections, and compare them with Eqs. (7)–(8) at q=2.2, f=0.9, s=1.0; if the true projections differ substantially, the closed-form mixing derivation is not the correct projection and the calibrated gains are artifacts of the synthetic reference.

Watch

Extended reading notes

Core claim

The central claim is that the tightest useful HBB for a ship is not the bounding box of the rotated rectangle but the projection of an inferred hull, and that a superellipse with q=2.2, f=0.9, and s=1.0 captures that hull well enough to improve conversion on real data. The paper derives the HBB half-extents by maximizing concave mixing functions over scalars u and v, obtaining t_x and t_y, and then applies a shrink factor. On ShipRSImageNet this yields a mean IoU of 0.5609, compared with 0.4929 for OuterHBB, 0.3139 for AreaEquivalentHBB, and 0.2451 for GBBMarginalized, while lowering overshoot to 0.3749 and undershoot to 0.1208; angle-stratified results show the gain holds in each orientation band from 0° to 45°. The authors frame the result as a calibrated rule, not a fixed default: synthetic experiments show the method needs parameter tuning before deployment.

Load-bearing premise

The load-bearing premise is that Eqs. (3)–(4) are the correct axis-aligned projection of the rotated superellipse in Eq. (2); the paper states this projection rather than deriving it from the boundary parameterization, and at θ=0 it returns x_h = b s and y_h = a s, swapping the ship's length and width.

Editorial extensions

If this is right

OBB-to-HBB conversion can be improved with a closed-form, parameterized rule, so detection pipelines can get tighter horizontal boxes without retraining the detector.
A single calibrated parameter set (q=2.2, f=0.9, s=1.0) transfers from synthetic hulls to real ship imagery, indicating synthetic calibration can stand in for costly manual tuning.
Tighter HBBs mean less empty background and fewer dropped pixels in crops, which reduces the data volume for downlink and improves shard quality for classifiers.
The angle-stratified results (IoU 0.7828, 0.4908, 0.3666 across 0°–15°, 15°–30°, 30°–45°) suggest the method avoids the sharp falloff that simple corner-enclosing conversions show for rotated ships.
For rectangular hulls the benefit is smaller, so the authors' stated conclusion is that the method must be calibrated for the target hull geometry before deployment.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The projection formulas in Eqs. (3)–(4) are stated rather than derived from the boundary parameterization in Eq. (2), so the chosen q=2.2 may partly compensate for a geometric mismatch; an exact projection of the superellipse could change the optimal parameters.
At θ=0 the formulas give x_h = b s and y_h = a s, swapping the ship's length and width; this suggests the coordinate convention for the input OBB must be checked before the rule is used elsewhere.
A natural extension, flagged by the paper's own rectangular-hull weakness, is to make q and f depend on aspect ratio L/W rather than using one global setting.
The same superellipse calibration could be tried for other elongated object classes such as aircraft or vehicles, but the parameters would likely need re-tuning per class.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this is a genuine new conversion rule — modeling a ship hull as a superellipse and tuning shape/fullness/shrink to convert OBB to HBB — but the paper's central projection formula is not the projection of that superellipse, and it swaps length and width at θ=0, contradicting its own definition of θ. The empirical gains are real but the stated mechanism is not.

What's good: The problem is practical. The paper ships code and data on GitHub/Zenodo, benchmarks on ShipRSImageNet and a Sentinel-2 set, defines overshoot/undershoot cleanly, and is honest that the method needs calibration and is sensitive to parameter choice. The idea of a shape-aware conversion with tuned hull parameters is new relative to the baselines cited.

Soft spots, in order of seriousness. First, Eqs. (3)-(4) are asserted as the axis-aligned projection but are not derived from Eq. (2). For a superellipse boundary, the y-coordinate at fixed u is b(1-u^q)^{1/q}, not b(1-u^q). The paper's degenerate case makes the error concrete: θ=0 gives half-extents (bs, as), swapping the ship's length and width even though θ is defined as the angle from the long axis. The correct support function is (a^p|cosθ|^p + b^p|sinθ|^p)^{1/p} with p=q/(q-1). So the grid search optimizes a misspecified geometric model, and the transfer of (q=2.2, f=0.9, s=1.0) to real data doesn't validate the superellipse mechanism.

Second, the headline synthetic IoU (0.7931) is in-sample: the same benchmark is used for the grid search and the reported number. Out-of-sample real-data results are better evidence, but the paper never shows synthetic IoU for the baselines, only relative area error, where NovelShapeAware is never best in any cell of Table I.

Third, the paper claims it 'simultaneously reduces overshoot and undershoot' relative to competing methods, but Table III shows OuterHBB has undershoot 0.0, and NovelShapeAware has 0.1208; on Sentinel-2 (Table IV) undershoot is 0.3562 vs OuterHBB's 0.2139. The claim is contradicted by the paper's own tables.

Minor: the method is called AreaEquivalentHBB in the text but 'M' is labeled 'Midpoint' in Table I, and no commit hash or baseline definitions are given for the code.

Bottom line: the paper is a promising calibration idea that needs a corrected projection formula and more careful reporting before the empirical comparison can be taken as evidence for the stated mechanism. I would send it to peer review because the idea is new, the data are real, and the flaw is fixable. But it is not citable in its current form.

Referee Report

3 major / 5 minor

Summary. The paper proposes a shape-aware OBB-to-HBB conversion method for ship detection, modeling the hull as a superellipse with parameters q (shape exponent), f (fullness), and s (shrink). It presents a projection formula (Eqs. 3-4), optimizes q, f, s on synthetic hulls via grid search, and evaluates the resulting converter on ShipRSImageNet and a custom Sentinel-2 dataset against OuterHBB, AreaEquivalentHBB, and GBBMarginalized. The authors report the highest mean IoU on both real datasets and argue that the superellipse model captures ship geometry better than box-corner or Gaussian baselines.

Significance. If correct, the method would provide a lightweight, retraining-free post-processing step for OBB-to-HBB conversion with tighter bounds. The paper ships code and data on GitHub/Zenodo, gives clear metric definitions, and includes out-of-sample real-data evaluation, which are commendable. However, the central geometric derivation is incorrect (see major comments), and the synthetic calibration is in-sample. As a result, the claimed mechanism is not established; what remains is an ad-hoc three-parameter heuristic whose empirical gains may be real but are not explained by the superellipse projection. The contribution in its current form does not meet the standard for publication.

major comments (3)

[III-B, Eqs. (3)-(4)] The half-extents in Eqs. (3)-(4) are not the projection of the rotated superellipse defined by Eq. (2). For a superellipse with exponent q, the true axis-aligned extent after rotation by theta is (a^p |cos theta|^p + b^p |sin theta|^p)^{1/p} with p = q/(q-1). The expression in Eq. (3) uses (1-u^q) instead of (1-u^q)^{1/q}, which changes the maximization. Concretely, at theta=0 Eq. (3) gives x_h = b s and Eq. (4) gives y_h = a s, swapping the ship's half-length and half-width, as noted in Section III-D. The correct values at theta=0 are x_h = a and y_h = b. Consequently, the optimal mixing parameters in Eqs. (5)-(6) and the final half-extents in Eqs. (7)-(8) maximize a misspecified objective. This invalidates the geometric grounding of the method and of the calibrated parameters.
[V-A and V-B] The grid search in Algorithm 2 selects q, f, and s using the same synthetic hull dataset on which the headline mean IoU (0.7931) and Match50 (0.9831) are reported. These numbers are therefore in-sample and do not provide evidence of generalization. Moreover, Table I shows that NovelShapeAware does not achieve the lowest relative error for any hull family or fullness level; AreaEquivalentHBB is best for tapered and fine hulls and OuterHBB for rectangular hulls. The paper should report a held-out or cross-validated synthetic evaluation, and should not present in-sample calibration performance as a result.
[V-D, Table III] The real-data evaluation on ShipRSImageNet is out-of-sample and provides the only independent support for the method. However, because the projection equations are wrong, the reported improvement could be an artifact of the three fitted parameters rather than evidence for the superellipse mechanism. The authors should either (a) derive the correct projection and re-run the calibration and real-data experiments, or (b) explicitly present the method as a heuristic and compare against other simple heuristics with the same number of parameters. Without this, the central claim that the converter is shape-aware in the geometric sense is unsupported.

minor comments (5)

[V-E] There is a typo in 'ccommonly used conversion methods' in Section V-E; it should be 'commonly'.
[III-A] The notation for half-width is inconsistent: the problem statement uses w, while Eq. (1) uses b; unify the symbols for clarity.
[Table I] Table I would benefit from a standard three-column layout with hull family and fullness as separate columns, rather than rows labeled as '0.60 - N'.
[Abstract] The sentence 'any current post-processing conversion from OBB to HBB either introduces excess empty and background space or removes data from the detection' is a fragment; it should be joined to the previous sentence or rewritten.
[V-D] The paper uses both 'AreaEquivalentHBB' and 'AreaEquivalenteHBB' (e.g., in Table III and Section V-D); please make the spelling consistent.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: synthetic calibration is in-sample, but real-data evaluation is out-of-sample and independent.

full rationale

The paper's derivation chain is not circular. The shape parameters (q, f, s) are calibrated on a synthetic hull dataset via Algorithm 2 and then applied unchanged to two out-of-sample real datasets (ShipRSImageNet and Sentinel-2), so the headline real-data IoU gains (0.5609 and 0.3527) are not fitted values. The synthetic mean IoU of 0.7931 is explicitly the calibration objective, not a prediction, and the paper labels the method a 'calibrated rule' (Section VIII), acknowledging the need for tuning. There is no load-bearing self-citation: the only references are external datasets and standard methods. The main caveat is geometric: Eqs. (3)-(4) are asserted as the projection of the rotated superellipse without derivation, and the θ≈0 case appears to swap half-length and half-width; this is a correctness risk, not a circular dependency, because the HBB is not defined in terms of the evaluation metric or fitted parameters by construction. The in-sample reporting of the synthetic benchmark is a methodological limitation but does not reduce the real-data claim to its inputs.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central algorithm depends on three fitted scalars, an unvalidated hull-shape assumption, and a projection formula that is asserted rather than derived. No new physical entities are introduced.

free parameters (3)

q (superellipse shape exponent) = 2.2 (grid search over {1.5, 1.8, 2.0, 2.2, 2.4, 2.6})
Chosen by mean IoU on the synthetic hull benchmark; controls how rectangular or pointed the hull model is.
f (fullness factor) = 0.9 (grid search over {0.7, 0.8, 0.9, 1.0, 1.1})
Scales the half-width b in the superellipse model; selected by mean IoU on the synthetic hull benchmark.
s (shrink factor) = 1.0 (grid search over {0.95, 1.0, 1.05, 1.1})
Scales the final HBB half-extents; selected by mean IoU on the synthetic hull benchmark.

assumptions (5)

domain assumption A ship hull cross-section is well approximated by a superellipse with parameters q and f.
Section III-A introduces the Lame curve model without independent validation against real hull geometry.
ad hoc to paper The maximum-over-mixing-scalars expression in Eqs. (3)-(4) equals the projection of the rotated superellipse onto the axis-aligned axes.
The expression is asserted in Section III-B; it is not derived from Eq. (2) and is inconsistent with the theta=0 case.
domain assumption Synthetic hull families are representative enough to calibrate q, f, and s for real ship detections.
Section V-B uses synthetic hulls to choose global parameters, then applies them to ShipRSImageNet and Sentinel-2.
domain assumption The annotated HBB in ShipRSImageNet and Sentinel-2 is the correct target for conversion.
IoU is computed against the dataset HBB; conversion quality is judged solely by agreement with this annotation.
domain assumption The input OBB is a minimum rotated rectangle fitted to the ship and the centroid is preserved in conversion.
Used throughout Algorithm 1; no validation of MRR accuracy is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Shape-Aware Oriented Bounding Box (OBB) to Horizontal Bounding Box (HBB) Conversion." pith.science (2026). https://pith.science/paper/GM6DDLJU

@misc{pith2026260805858,
  author       = {Pith},
  title        = {Pith review of: Shape-Aware Oriented Bounding Box (OBB) to Horizontal Bounding Box (HBB) Conversion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GM6DDLJU}},
  note         = {Machine review of arXiv:2608.05858}
}

read the original abstract

Accurate object detection in aerial and satellite imagery is dependent upon the bounding box representation. This is especially true for spatially oriented objects such as ships or aircrafts. Oriented Bounding Boxes (OBB) have a tighter fit and more robust non-max suppression compared to Horizontal Bounding Boxes (HBB), any current post-processing conversion from OBB to HBB either introduces excess empty and background space or removes data from the detection. This paper introduces a novel approach for a shape-aware OBB-to-HBB conversion for ship detection in remote sensing imagery. It leverages hull shape, hull fullness, and the bounding box orientation to produce a tighter axis-aligned HBB representation. The proposed method is benchmarked against three baselines methods for OBBto-HBB conversion, Outer HBB which uses minimum and maximum, Area Equivalent HBB and GBB Marginalized HBB.

Figures

Figures reproduced from arXiv: 2608.05858 by the authors.

**Figure 3.** Best relative error achieved by the shape-aware [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Angle-stratified mean IoU on ShipRSImageNet [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

Figures from the paper (1 more)

**Figure 6.** Figure 6: Showcasing NovelShapeAware HBB (yellow) with Ground truth for OBB (Red) and HBB (Green) [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

10 extracted references · 7 canonical work pages

[1]

Faster r-cnn: Towards real-time object detection with region proposal networks,

S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-time object detection with region proposal networks,” 2015

work page 2015
[2]

Cornernet: Detecting objects as paired keypoints,

H. Law and J. Deng, “Cornernet: Detecting objects as paired keypoints,”CoRR, vol. abs/1808.01244, 2018. [Online]. Available: http://arxiv.org/abs/1808.01244

arXiv 2018
[4]

Arbitrary-Oriented Scene Text Detection via Rotation Proposals

J. Ma, W. Shao, H. Ye, L. Wang, H. Wang, Y . Zheng, and X. Xue, “Arbitrary-oriented scene text detection via rotation proposals,”CoRR, vol. abs/1703.01086, 2017. [Online]. Available: http://arxiv.org/abs/1703.01086

work page Pith review arXiv 2017
[5]

Learning non-maximum suppression

J. H. Hosang, R. Benenson, and B. Schiele, “Learning non-maximum suppression,”CoRR, vol. abs/1705.02950, 2017. [Online]. Available: http://arxiv.org/abs/1705.02950

work page Pith review arXiv 2017
[6]

Dual- det : a fast detector for oriented object detection in aerial images,

Q. Guan, Z. Qu, P. Zhao, M. Zeng, and J. Liu, “Dual- det : a fast detector for oriented object detection in aerial images,”International Journal of Remote Sensing, vol. 42, no. 24, pp. 9542–9564, 2021. [Online]. Available: https://doi.org/10.1080/01431161.2021.1995071

arXiv 2021
[7]

Oriented bounding box detection algorithm for dense scenarios of robotic arm operation,

J. Dong, L. Deng, D. Wan, C. Liu, J. Yin, M. Guo, H. Zhang, S. Lin, H. Liu, and L. Liu, “Oriented bounding box detection algorithm for dense scenarios of robotic arm operation,”Expert Systems with Applications, vol. 298, p. 129678, 2026. [Online]. Available: https://www.sciencedirect. com/science/article/pii/S0957417425032932

work page 2026
[8]

BoxInst: High-Performance Instance Segmentation with Box Annotations

Z. Tian, C. Shen, X. Wang, and H. Chen, “Boxinst: High- performance instance segmentation with box annotations,” CoRR, vol. abs/2012.02310, 2020. [Online]. Available: https: //arxiv.org/abs/2012.02310

work page Pith review arXiv 2012
[9]

An automated method for the creation of oriented bounding boxes in remote sensing ship detection datasets,

G. Savathrakis and A. Argyros, “An automated method for the creation of oriented bounding boxes in remote sensing ship detection datasets,” in2024 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), 2024, pp. 830–839

work page 2024

Show all 10 references

[10]

ShipRSIma- geNet: A Large-Scale Fine-Grained Dataset for Ship Detection in High-Resolution Optical Remote Sensing Images,

Z. Zhang, L. Zhang, Y . Wang, P. Feng, and R. He, “ShipRSIma- geNet: A Large-Scale Fine-Grained Dataset for Ship Detection in High-Resolution Optical Remote Sensing Images,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 14, pp. 8458–8472...

2021
[2017]

Available: http://arxiv.org/abs/1711.10398

[Online]. Available: http://arxiv.org/abs/1711.10398

arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Faster r-cnn: Towards real-time object detection with region proposal networks,

S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-time object detection with region proposal networks,” 2015

work page 2015

[2] [2]

Cornernet: Detecting objects as paired keypoints,

H. Law and J. Deng, “Cornernet: Detecting objects as paired keypoints,”CoRR, vol. abs/1808.01244, 2018. [Online]. Available: http://arxiv.org/abs/1808.01244

arXiv 2018

[3] [4]

Arbitrary-Oriented Scene Text Detection via Rotation Proposals

J. Ma, W. Shao, H. Ye, L. Wang, H. Wang, Y . Zheng, and X. Xue, “Arbitrary-oriented scene text detection via rotation proposals,”CoRR, vol. abs/1703.01086, 2017. [Online]. Available: http://arxiv.org/abs/1703.01086

work page Pith review arXiv 2017

[4] [5]

Learning non-maximum suppression

J. H. Hosang, R. Benenson, and B. Schiele, “Learning non-maximum suppression,”CoRR, vol. abs/1705.02950, 2017. [Online]. Available: http://arxiv.org/abs/1705.02950

work page Pith review arXiv 2017

[5] [6]

Dual- det : a fast detector for oriented object detection in aerial images,

Q. Guan, Z. Qu, P. Zhao, M. Zeng, and J. Liu, “Dual- det : a fast detector for oriented object detection in aerial images,”International Journal of Remote Sensing, vol. 42, no. 24, pp. 9542–9564, 2021. [Online]. Available: https://doi.org/10.1080/01431161.2021.1995071

arXiv 2021

[6] [7]

Oriented bounding box detection algorithm for dense scenarios of robotic arm operation,

J. Dong, L. Deng, D. Wan, C. Liu, J. Yin, M. Guo, H. Zhang, S. Lin, H. Liu, and L. Liu, “Oriented bounding box detection algorithm for dense scenarios of robotic arm operation,”Expert Systems with Applications, vol. 298, p. 129678, 2026. [Online]. Available: https://www.sciencedirect. com/science/article/pii/S0957417425032932

work page 2026

[7] [8]

BoxInst: High-Performance Instance Segmentation with Box Annotations

Z. Tian, C. Shen, X. Wang, and H. Chen, “Boxinst: High- performance instance segmentation with box annotations,” CoRR, vol. abs/2012.02310, 2020. [Online]. Available: https: //arxiv.org/abs/2012.02310

work page Pith review arXiv 2012

[8] [9]

An automated method for the creation of oriented bounding boxes in remote sensing ship detection datasets,

G. Savathrakis and A. Argyros, “An automated method for the creation of oriented bounding boxes in remote sensing ship detection datasets,” in2024 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), 2024, pp. 830–839

work page 2024