Object Distortion Control (O-DisCo)

• Object Distortion Control (O-DisCo) is a framework that explicitly represents and manages distortions affecting object geometry, appearance, or structure.
• It integrates both physics-driven calibration and learned control mechanisms to accurately correct hardware-induced, refractive, and projection distortions.
• O-DisCo is applied in diverse domains such as OCT, refractive vision, panoramic imaging, video editing, and graph-based structural preservation.

Searching arXiv for recent and foundational papers relevant to O-DisCo and related distortion-control formulations. Object Distortion Control (O-DisCo) denotes a family of formulations in which distortion affecting object geometry, appearance, or retained pairwise structure is explicitly represented and then corrected, constrained, or intentionally modulated. In the cited literature, the term appears both as a broad cross-domain viewpoint and as a concrete control signal in unified realistic video editing. Under this reading, O-DisCo spans calibration-based correction of scanner-induced fan distortion in OCT, explicit refractive-surface estimation for cameras behind transparent media, ray-geometry compensation in tomographic additive manufacturing, controllable fisheye rectification, distortion-aware panoramic depth estimation, self-supervised mesh-based retargeting, sequential algorithm selection for corruption removal, object-region adversarial perturbation with distortion budgets, and even distortion-bounded removal of nonterminal graph objects (Jeught et al., 2012, Pável et al., 2019, Orth et al., 2021, Guo et al., 2024, Chen et al., 2020, Shen et al., 2022, Liao et al., 31 Oct 2025, Kapoor et al., 2024, Phuc et al., 2024, Filtser, 2017, Chen et al., 1 Sep 2025).

1. Conceptual scope

Across these works, O-DisCo is not a single algorithmic primitive. It is better understood as a control paradigm in which the object-relevant consequences of distortion are elevated to first-class variables. The distortion may be hardware-induced, as in scanner/lens geometry in wide-field OCT; interface-induced, as in refractive covers or cylindrical vials; projection-induced, as in fisheye or equirectangular imagery; representation-induced, as in cubemap seam discontinuities; or intentionally injected as an editable control signal in generative video models (Jeught et al., 2012, Pável et al., 2019, Orth et al., 2021, Guo et al., 2024, Shen et al., 2022, Chen et al., 1 Sep 2025).

A second unifying feature is that O-DisCo is object-centric even when the mechanism is global. In OCT and TAM, the target is faithful recovery of object surface shape or printed object geometry. In panoramic depth and retargeting, the goal is to preserve object structure despite spatially varying projection distortion. In DeepClean, the objective is to identify the current corruption and select a corrective operator that improves downstream object detection. In Steiner point removal, the “objects” are terminals whose pairwise geometry is retained while all Steiner objects are removed with bounded distortion (Jeught et al., 2012, Orth et al., 2021, Chen et al., 2020, Liao et al., 31 Oct 2025, Kapoor et al., 2024, Filtser, 2017).

Domain	Representative formulation	O-DisCo role
Large-volume OCT	3D recalibration via successive radial remappings	Real-time correction of scanner/lens distortion
Refractive vision	Explicit ray tracing through a modeled refractive shell	Depth-aware distortion estimation
Fisheye/panoramas	Query-conditioned rectification or distortion-aware features	Learned spatial control of rectification
Video editing	R-O-DisCo and A-O-DisCo	Unified object-region control signal
Image retargeting	Object-aware mesh deformation	Preserve important object geometry
Graph compression	Steiner point removal	Bound distortion under object removal

This suggests that O-DisCo is defined less by modality than by a recurrent technical question: which representation of distortion is sufficiently structured to preserve or steer object fidelity under computational and data constraints?

2. Distortion representations and control variables

The literature recurrently replaces raw image-space heuristics with explicit distortion variables. In OCT, the distortion is represented by virtual radial centers and scan radii $R_x$ and $R_y$, estimated from a flat mirror and used in successive polar-to-Cartesian remappings of $D(x,y,z)$ to corrected coordinates $(x',y',z'')$ (Jeught et al., 2012). In blind geometric correction, the representation is a dense forward flow $f(p)$ such that $q=p+f(p)$, optionally followed by parametric fitting through a distortion model $\mathcal{M}^{\beta}$ and Hough voting in parameter space (Li et al., 2019).

In refractive calibration, the control variable is the geometry of the refractive interface itself. The forward model is $$f_{\theta}(\mathbf{p}) : \Omega \rightarrow \mathbb{R}^3$$, with an outer-surface irregularity parameterized by an RBF network,

$$\Phi(\mathbf{s}^\prime) = \sum_{i,j=1}^{N} a_{ij}\,\mathrm{RBF}(\mathbf{s}_{ij},\mathbf{s}^\prime),$$

and estimated by minimizing checkerboard reprojection error through differentiable ray tracing (Pável et al., 2019). In TAM, the central representation is a ray-coordinate remapping from real projector coordinates to virtual parallel-beam coordinates,

$$(x_p,\theta)\rightarrow \big(x_v(x_p),\theta_v(x_p,\theta)\big),$$

augmented by Jacobian and Fresnel weighting of the resampled sinogram (Orth et al., 2021).

Learned O-DisCo variants replace physical parameters with spatially structured latent controls. QueryCDR associates each distortion degree $R_y$0 with a learnable query $R_y$1, propagates it as layer-wise control conditions $R_y$2, and uses these controls in CCMB and CAMB blocks. The queries can also be interpolated continuously, for example $R_y$3 and $R_y$4 (Guo et al., 2024). O-DisCo-Edit makes this idea explicit in the name itself: R-O-DisCo is a masked stochastic distortion of the editable object region during training, while A-O-DisCo is an adaptive, contrast-scaled and Gaussian-blurred distortion signal used at inference,

$R_y$5

This turns distortion from nuisance to control channel (Chen et al., 1 Sep 2025).

The common implication is that O-DisCo systems are organized around an intermediate distortion representation rather than a direct “distorted $R_y$6 corrected” black box. The representation may be geometric, optical, latent, or mesh-based, but it is always intended to make object-level consequences controllable.

3. Calibration- and physics-driven O-DisCo

A major branch of O-DisCo is empirical or physics-based correction of deterministic acquisition geometry. The OCT recalibration method in “Large-volume optical coherence tomography with real-time correction of geometric distortion artifacts” treats scanner-induced fan distortion as predominantly radial in each scan direction, calibrates it with a flat reflective mirror, and applies two successive 2D remappings followed by 3D resampling on a GPU. In a $R_y$7 volume, the method reduced the root-mean-square mismatch of a human tympanic membrane surface from $R_y$8 to $R_y$9, while uncorrected edge errors of almost $D(x,y,z)$0 in a central cross-section were reduced to about $D(x,y,z)$1 across the full B-scan length. The reported implementation achieved real-time interpolation on an NVIDIA GTX570 GPU, and the paper emphasized that transfer cost, not interpolation cost, dominated the runtime (Jeught et al., 2012).

The refractive-camera work “Distortion Estimation Through Explicit Modeling of the Refractive Surface” addresses a different distortion class: a camera imaging through a refractive shell for which a single central projection ray is invalid. Here O-DisCo takes the form of explicit geometry identification. Rays intersect an inner cone, refract through the medium according to Snell’s law, intersect an outer cone with learned irregularity, refract again, and finally hit a calibration plane. The inverse problem estimates the RBF amplitudes $D(x,y,z)$2 by minimizing

$D(x,y,z)$3

A central result is that the distortion vector $D(x,y,z)$4 depends on depth and varies approximately linearly on inverse depth, which means the distortion is not a single image-plane warp independent of scene geometry (Pável et al., 2019).

“Correcting ray distortion in tomographic additive manufacturing” shows a related but distinct physics-driven O-DisCo strategy. Instead of forcing the optical hardware to approximate ideal parallel-beam tomography through an index-matching bath or cylindrical lens, it remaps the projection data itself by resampling the Radon transform into the true aberrated geometry. The method models both cylindrical-vial lensing and non-telecentricity, computes $D(x,y,z)$5 for each real projector coordinate $D(x,y,z)$6, and applies Jacobian and Fresnel compensation before projection. The paper states that correcting both lensing and non-telecentricity restored the dose field close to the target, while hardware index matching alone did not eliminate non-telecentric distortion (Orth et al., 2021).

A common misconception is that physically faithful O-DisCo always requires full ray tracing through the sample or scene. The OCT paper explicitly positions empirical recalibration as an alternative to ray tracing for real-time external surface recovery, while the TAM paper shows that geometric optics can often be handled by pre-distorting the control signal rather than redesigning the hardware (Jeught et al., 2012, Orth et al., 2021). Conversely, the refractive-camera paper shows why a single precomputed image warp is insufficient when distortion is depth-dependent (Pável et al., 2019).

4. Learned distortion-aware rectification and dense prediction

Another branch of O-DisCo learns distortion-aware representations directly in the network. QueryCDR formulates fisheye rectification as a controllable process rather than a fixed mapping. Its Distortion-aware Learnable Query Mechanism extracts layer-specific control conditions $D(x,y,z)$7 from a user-selected query $D(x,y,z)$8, and these conditions modulate a U-shaped hybrid network through CCMB in layers $D(x,y,z)$9 and CAMB in layers $(x',y',z'')$0. The controllability is discrete when a single query is selected and continuous when neighboring queries are interpolated. On the COCO synthetic fisheye dataset with 9 distortion degrees, QueryCDR achieved average PSNR $(x',y',z'')$1 and SSIM $(x',y',z'')$2, compared with the second-best SimFIR at PSNR $(x',y',z'')$3 and SSIM $(x',y',z'')$4. Its ablations also show that learnable spatial control outperformed scalar and fixed-query alternatives, with average PSNR $(x',y',z'')$5 for “Learnable Query” versus $(x',y',z'')$6 for “Scalar” and $(x',y',z'')$7 for “W/o Control” (Guo et al., 2024).

DAMO addresses equirectangular panorama distortion in dense depth estimation by combining three mechanisms: deformable convolution for adaptive local sampling, a Strip Pooling Module for anisotropic context aggregation, and a spherical-aware weight matrix $(x',y',z'')$8 for correcting supervision bias caused by unequal spherical area representation. The final loss is

$(x',y',z'')$9

where $f(p)$0 is BerHu loss. On the 360D dataset, the full model reported RMSE $f(p)$1, Abs$f(p)$2 $f(p)$3, RMSE(log) $f(p)$4, $f(p)$5, $f(p)$6, and $f(p)$7. The ablations attribute gains to both adaptive local geometry and projection-aware supervision: deformable convolution improved over the base model with only $f(p)$8M parameters, while SPM yielded the largest single architectural gain (Chen et al., 2020).

The dual-cubemap method for omnidirectional depth estimation takes a representational rather than feature-only approach. It shifts depth estimation from a highly distorted equirectangular domain to two complementary cubemap decompositions, one obtained after a $f(p)$9 horizontal rotation. The DCDE module performs depth estimation on both cubemaps, and the BR module then resolves seam-induced discontinuities in the reconstructed equirectangular depth through an encoder-decoder trained with Berhu and gradient losses. On Stanford2D3D, a single-cubemap-only setup reported MAE $q=p+f(p)$0 and $q=p+f(p)$1, whereas the full dual-cubemap + BR + GL system achieved MAE $q=p+f(p)$2 and $q=p+f(p)$3 (Shen et al., 2022).

These systems illustrate two distinct learned O-DisCo philosophies. QueryCDR treats distortion as a controllable latent condition field; DAMO treats it as a mismatch between fixed CNN assumptions and spherical image geometry; dual-cubemap methods relocate inference into lower-distortion coordinates and repair the seams afterward. This suggests that learned O-DisCo is not synonymous with purely model-free rectification: geometry can enter through control tensors, sampling operators, projection-aware losses, or a change of representation itself (Guo et al., 2024, Chen et al., 2020, Shen et al., 2022).

5. Object-centric O-DisCo in editing and retargeting

“O-DisCo-Edit: Object Distortion Control for Unified Realistic Video Editing” gives the most literal definition of O-DisCo. It introduces Object Distortion Control as a unified signal for first-frame-guided video editing, replacing task-specific conditions such as masks, bounding boxes, optical flow, and tracking points with a single distortion-based representation localized to the edited region. R-O-DisCo is generated during training by random color distortion, mosaicking, and masked compositing,

$q=p+f(p)$4

whereas A-O-DisCo uses adaptive contrast scaling and Gaussian blur,

$q=p+f(p)$5

The framework combines O-DisCo with copy-form preservation,

$q=p+f(p)$6

and identity preservation on a CogVideoX-I2V backbone initialized from Diffusion as Shader. It supports object removal, outpainting, object internal motion transfer, lighting transfer, color change, swap, addition, and style transfer. The reported training regime used approximately 180k video-mask pairs, two LoRAs, and $q=p+f(p)$7K total steps, while the human evaluations favored O-DisCo-Edit across nearly all tasks except color change (Chen et al., 1 Sep 2025).

Object-IR specializes O-DisCo to content-aware image retargeting. It reformulates retargeting as learned mesh deformation, predicts the motion of each mesh grid vertex, and constrains the deformation of important object regions through three losses: object-consistent loss,

$q=p+f(p)$8

geometric-preserving loss on important mesh edges,

$q=p+f(p)$9

and boundary loss. The importance map is detector-driven: YOLO11 bounding boxes define the object regions that receive geometric protection. On RetargetMe, the method received $\mathcal{M}^{\beta}$0 pairwise votes, or $\mathcal{M}^{\beta}$1, and the abstract reports average inference time $\mathcal{M}^{\beta}$2 s for $\mathcal{M}^{\beta}$3 resolution (Liao et al., 31 Oct 2025).

Taken together, these two papers make an important distinction. In O-DisCo-Edit, distortion is a generative condition intentionally injected to steer synthesis. In Object-IR, distortion is a deformation that must be allocated away from semantically important regions. The same term therefore covers both controlled distortion injection and controlled distortion suppression, provided the object-level target is explicit (Chen et al., 1 Sep 2025, Liao et al., 31 Oct 2025).

6. Sequential selection, adversarial injection, and structural distortion bounds

O-DisCo also appears as a control policy over which corrective action to apply. DeepClean models image restoration as a hierarchical sequential decision problem with

$\mathcal{M}^{\beta}$4

where the high-level module predicts the latest corruption class and the low-level module chooses a corrective algorithm by minimizing cosine similarity in a corruption-aware embedding space,

$\mathcal{M}^{\beta}$5

The system is queried iteratively until the image is predicted as clean. On the COCO-based benchmark, DeepClean reached test mAP $\mathcal{M}^{\beta}$6 versus oracle HC-MTL at $\mathcal{M}^{\beta}$7, corresponding to $\mathcal{M}^{\beta}$8 normalized mAP, and substantially exceeded hard-coded and random baselines (Kapoor et al., 2024).

The adversarial-attack work “Distortion-Aware Adversarial Attacks on Bounding Boxes of Object Detectors” demonstrates the inverse use of O-DisCo: distortion is added rather than removed, but its amount and spatial support are explicitly controlled. The perturbation is restricted to the current union of predicted bounding boxes,

$\mathcal{M}^{\beta}$9

with total detector loss $$f_{\theta}(\mathbf{p}) : \Omega \rightarrow \mathbb{R}^3$$0. Distortion is monitored using

$$f_{\theta}(\mathbf{p}) : \Omega \rightarrow \mathbb{R}^3$$1

and the attack stops once a target distortion $$f_{\theta}(\mathbf{p}) : \Omega \rightarrow \mathbb{R}^3$$2 or desired success rate $$f_{\theta}(\mathbf{p}) : \Omega \rightarrow \mathbb{R}^3$$3 is reached. The paper reports up to $$f_{\theta}(\mathbf{p}) : \Omega \rightarrow \mathbb{R}^3$$4 success in white-box attacks and up to $$f_{\theta}(\mathbf{p}) : \Omega \rightarrow \mathbb{R}^3$$5 in black-box attacks, while arguing that object-region masking keeps perturbations more localized and more visually controlled (Phuc et al., 2024).

A more abstract form of O-DisCo appears in Steiner point removal. Given a weighted graph $$f_{\theta}(\mathbf{p}) : \Omega \rightarrow \mathbb{R}^3$$6 and a terminal set $$f_{\theta}(\mathbf{p}) : \Omega \rightarrow \mathbb{R}^3$$7, the goal is to remove all Steiner vertices while preserving terminal-terminal distances within multiplicative distortion. The paper proves that the ball-growing algorithm returns a minor $$f_{\theta}(\mathbf{p}) : \Omega \rightarrow \mathbb{R}^3$$8 with distortion $$f_{\theta}(\mathbf{p}) : \Omega \rightarrow \mathbb{R}^3$$9 with probability $$\Phi(\mathbf{s}^\prime) = \sum_{i,j=1}^{N} a_{ij}\,\mathrm{RBF}(\mathbf{s}_{ij},\mathbf{s}^\prime),$$0,

$$\Phi(\mathbf{s}^\prime) = \sum_{i,j=1}^{N} a_{ij}\,\mathrm{RBF}(\mathbf{s}_{ij},\mathbf{s}^\prime),$$1

This is a structural rather than optical O-DisCo: designated objects are retained, removable objects are eliminated, and pairwise geometry among the retained objects is controlled (Filtser, 2017).

These three cases show that O-DisCo can operate at different levels of abstraction: operator selection, perturbation budgeting, and combinatorial structure preservation. The commonality is explicit control over how much object-relevant geometry or recognizability is allowed to change.

7. Limitations, misconceptions, and research directions

A recurring misconception is that distortion correction can always be collapsed into a single image-plane warp. The refractive-camera paper explicitly shows that distortion depends on depth through $$\Phi(\mathbf{s}^\prime) = \sum_{i,j=1}^{N} a_{ij}\,\mathrm{RBF}(\mathbf{s}_{ij},\mathbf{s}^\prime),$$2, and the TAM paper shows that the effective projection angle inside the medium depends jointly on projector coordinate and rotation angle. This means that a single depth-independent undistortion is insufficient whenever refractive or non-telecentric effects dominate (Pável et al., 2019, Orth et al., 2021).

A second misconception is that successful O-DisCo must always be either fully physical or fully learned. The literature instead supports a continuum. OCT uses empirical calibration and lookup tables for real-time correction; refractive calibration uses explicit differentiable ray tracing with an RBF surface prior; QueryCDR uses spatially dense learnable queries without explicit camera-parameter grounding; DAMO inserts geometry through adaptive sampling and spherical weighting; Object-IR uses detector-derived object boxes and mesh deformation; O-DisCo-Edit uses a learned distortion signal in latent diffusion space (Jeught et al., 2012, Guo et al., 2024, Chen et al., 2020, Liao et al., 31 Oct 2025, Chen et al., 1 Sep 2025).

The limitations are correspondingly heterogeneous. OCT correction is setup-specific and does not fully recover true internal 3D anatomy in refractive or multilayered objects without additional processing. Explicit refractive-surface modeling assumes a fixed base shape and a controlled checkerboard calibration procedure, and is therefore more naturally offline than real-time. QueryCDR models nine distortion degrees and demonstrates interpolation across trained degrees, but does not establish out-of-range extrapolation. DeepClean remains global rather than object-aware. O-DisCo-Edit depends strongly on the quality of the edited first frame and reports failure on complex four-limbed animal swap. Object-IR weakens when salient objects and line structures are scattered across the image or when even slight human-body distortion becomes conspicuous (Jeught et al., 2012, Pável et al., 2019, Guo et al., 2024, Kapoor et al., 2024, Chen et al., 1 Sep 2025, Liao et al., 31 Oct 2025).

A plausible implication is that future O-DisCo systems will be layered. The existing literature already suggests one decomposition: first remove deterministic hardware- or representation-induced distortion with setup-specific calibration or geometry-aware representation changes; then address object-dependent residuals with learned spatial control, mesh constraints, or generative priors; and finally choose or adapt operators sequentially when the distortion class is not known a priori (Jeught et al., 2012, Shen et al., 2022, Kapoor et al., 2024). Another plausible implication is that object-level distortion control will increasingly require hybrid conditioning: explicit geometry where physics is stable, latent spatial control where distortion is diverse, and region-preserving mechanisms where edits or retargeting are localized.

In this broader sense, O-DisCo designates a technical program rather than a narrow subfield: isolate distortion as a structured variable, couple that variable to object fidelity, and then optimize either correction, preservation, or manipulation under the constraints imposed by the sensing, rendering, or compression process.