PaperlayerView exact revision on arXiv

Source-linked AI summary

Selective Monocular Precession Estimation under Geometric and Evaluation Ambiguity: An Auditable Synthetic Study with ORBIT-SCOUT

Carlos Eduardo Sanoja

arXiv:2610.34819v1cs.CV

TL;DR

Monocular images can yield a precise fitted trajectory without establishing that its recovered axis is physically correct. The paper evaluates a selective, inspectable precession-estimation implementation on new synthetic scenes and reports correct estimates alongside substantial withholding and unmet registered targets. Its findings are bounded by shared image-feature uncertainty and the small synthetic evaluation.

  • Problem

    A fitted monocular trajectory may not establish that the recovered physical axis is correct, because projected lines, rims and texture can be ambiguous.

    Source: §1 · p. 1

  • Method

    The paper evaluates a selective, inspectable monocular precession implementation and separates fresh-population evaluation from development replays and repeated scenes.

    Source: §2.7 · p. 7, §9 · p. 20

  • Results

    The primary pass produced 38 correct joint estimates, no wrong emissions, 57 withholdings and one incomplete execution; registered coverage and complete-execution requirements were not met.

    Source: §6.1 · p. 13, §9 · p. 20

  • Takeaways & Limitations

    The observed result is a narrow synthetic case-study finding, not a universal identifiability claim or a ranking against methods evaluated on different observables.

    Source: §1 · p. 2

    Only 96 scenes from four assemblies were drawn, and no calibrated probability model for accepted error was validated, so conditional-error summaries should not be extrapolated to deployment.

    Source: §8.3 · p. 19

Abstract

from arXiv · show

Monocular observations of a tumbling object can support a precise fitted trajectory without establishing that the recovered physical axis is correct. We study this distinction for elongated, approximately cylindrical targets observed by a calibrated stationary camera. ORBIT-SCOUT couples image-derived line and rim geometry to constant-cone and torque-free triaxial trajectory families, then returns a linked momentum-axis and mean-precession-rate estimate only when numerical, geometric, temporal and competing-hypothesis checks pass. We describe the measurement objective, selection policy and limitations, and situate them within spacecraft pose estimation, unknown-target reconstruction, rigid-body identification and selective prediction. An internally preregistered population of 96 new synthetic scenes was processed in two cold passes, retaining all 192 attempts. In the primary pass, 38 scenes produced jointly correct estimates, none produced an incorrect estimate, 57 were withheld and one execution was incomplete. Correct coverage on the 64 supported scenes was 25/64. Ninety-five repeated pairs were scientifically equivalent under a frozen comparison policy; the remaining pair was incomplete. The registered coverage, failure-free execution and all-case repeatability targets were not met. Historical audits also exposed an overly permissive axis metric and motion-dependent acquisition cadence. The contribution is a documented measurement-to-decision implementation and a failure-preserving evaluation case study, not a new rigid-body theory or evidence of state-of-the-art accuracy. Zero observed errors does not establish calibrated risk, and conditional temporal checks do not remove shared image-feature bias. Real-image transfer and external reproduction remain open.

1 Introduction

The paper frames monocular tumbling-object estimation as a question of whether images support the intended physical quantity, not merely whether a trajectory fits. It documents a selective implementation and a problem-oriented review, while keeping the synthetic evaluation’s failures and comparison limits explicit.

1 Introduction

  • ORBIT-SCOUT documents a geometry-to-decision implementation; its 96-scene primary pass emitted 38 correct joint estimates, no observed incorrect emissions, and substantial abstention, failing registered validation.The paper presents this as a bounded case study, not a safety probability or ranking against methods using different observables.Source: §1 · p. 2

2.1 Review procedure

  • The extended problem-oriented review is not exhaustive, and numerical scores are not pooled across incompatible tasks.It prefers primary sources and limits claims about unavailable full texts to accessible abstracts and previews.Source: §2.1 · p. 3

2.2 Known-target monocular pose estimation

  • Known-target pose methods use CAD or other target information, so their pose benchmarks do not directly rank an unoriented momentum-axis and precession-rate estimator.The review distinguishes known-target systems from unknown-target settings and notes that adapting full-orientation outputs requires a defined temporal estimator and reference.Source: §2.2 · p. 3

2.3 Temporal pose tracking and dynamics

  • Temporal consistency can enter at different pipeline stages; ORBIT-SCOUT’s checks follow shared image measurements, and its finite offline rival bank is a veto rather than a global posterior.The review contrasts learned pose tracking, retained pose histories, and temporal representation methods without claiming a direct performance comparison.Source: §2.3 · p. 4

2.4 Unknown-target reconstruction and navigation

  • Unknown-target methods vary in auxiliary sensing and reconstruction priors, so a single camera does not specify the information available to an estimator.Examples include auxiliary attitude information, learned shape priors, and target-specific reconstruction; reducing one prior can be offset by another.Source: §2.4 · pp. 4–5

2.5 Rotational state, inertia and circular geometry

  • Chi et al. fit symmetric-target rotation, Zhang et al. [6] combine circular features with dynamics, and torque-free rigid-body solutions are established mathematics.The paper locates its validation need in connecting extracted image structure to the physical trajectory.Source: §2.5 · pp. 5–6

2.6 Datasets and the simulation-to-reality gap

  • The all-synthetic renderer varies geometry, viewing direction, lighting and material, but omits sensor noise and motion blur and includes no real or testbed images.Thus appearance transfer and operational sensor-condition performance are not measured.Source: §2.6 · p. 6

2.7 Selective prediction, uncertainty and evaluation integrity

  • Fixed engineering gates allow withholding but do not inherit statistical guarantees; calibrated confidence is absent, and policy selection must be separated from fresh evaluation.The study separates development replays, a fresh population and repeated runs on the same scenes.Source: §2.7 · pp. 6–7

3 Problem, estimand and physical assumptions

The problem is restricted to estimating a camera-frame, unoriented angular-momentum axis and positive mean precession rate from calibrated monocular video. The estimand and supported physical assumptions exclude several broader pose, motion and inertia quantities.

Inputs and model assumptions

  • Inference takes calibrated images and their acquisition calendar; the supported model assumes a stationary camera and alignment of the optical longitudinal axis with a principal inertia axis.Target attitude, inertia, momentum, stratum labels and reference motion are not inference inputs.Source: §3 · p. 7

Estimand

  • The requested axis is the unoriented camera-frame angular-momentum line, and mean precession is the least-squares slope of unwrapped phase across the full acquisition calendar.It is not an endpoint secant, instantaneous rate, or a slope restricted to frames admitted by feature extraction; a nonzero visual frequency is not automatically this physical mean.Source: §3 · p. 7

Decision outcomes

  • A withheld decision means checks did not support a point estimate, whereas incomplete execution means the decision could not be established; neither counts as correct.Withholding does not by itself establish physical unobservability.Source: §3 · p. 8

Physical assumptions

  • The supported setting assumes a rigid, approximately cylindrical target, accurate times and calibration, fixed range, and negligible external torque; absolute inertia and metric size are not inferred.Noncylindrical shape, optical/inertial misalignment and biased calibration are stress conditions rather than part of the supported domain.Source: §3 · p. 8

4 Measurement-to-decision method

The method connects image-derived geometry to two rigid-body trajectory families, then reports estimates only after fixed selection and acceptance checks. Its temporal and geometric checks remain conditional on the extracted features and finite search.

4.1 Image measurements and their conditional nature

  • Line groups and rim candidates are measured, but feature identity remains unresolved because whole-video preprocessing makes later temporal checks conditional on shared features.Source: §4.1 · p. 8

4.2 Endpoint projection and positive-ray support

  • The endpoint objective is a pixel-space correction with positive-ray support gates and a robust minimum across competing groups, not a demonstrated likelihood.Source: §4.2 · p. 9

4.3 Constant-cone and torque-free triaxial families

  • The estimator searches physically constrained constant-cone and torque-free triaxial families with finite prescribed starts and fixed bounds, including a 30 degrees/s angular-scale ceiling.A complete bank means all prescribed starts were retained, not that all physical solutions were found.Source: §4.3 · pp. 9–10

4.4 Fixed family selection and separate acceptance

  • Family selection uses a frozen dimension-penalized engineering rule; acceptance cannot substitute a runner-up when the selected fit fails.The statistical derivation of a Bayesian information criterion does not automatically apply to the correlated frames and robust OR loss.Source: §4.4 · p. 10

4.5 Temporal checks and competing trajectories

  • Emission requires all 27 gates, including eight completed temporal blocks and rival checks; reported radii summarize sensitivity, not confidence regions.Source: §4.5 · pp. 10–11

5 Evaluation design and provenance

The evaluation fixes a synthetic population and retains every attempt across two full cold passes, while scoring linked axis-and-rate correctness separately from withholding and incomplete execution. The authenticated descriptive supplement preserves the original failures but does not complete the planned same-population comparison.

5.1 New synthetic population

  • The fixed-seed population has 96 scenes (32 axisymmetric, 16 CN, 16 DN, 32 stress), with the first 64 designated supported and four recurring optical assemblies.Source: §5.1 · p. 11

5.2 Isolation, compute and repetitions

  • Two cold passes recomputed the full image-to-report pipeline; all 192 attempts stayed in the denominator without replacement under internal, not external, registration.Source: §5.2 · p. 12

5.3 Linked joint metric and outcome accounting

  • Joint success requires one emitted estimate to meet both axis <10° and rate <15% criteria; accounting distinguishes correct, wrong, withheld and incomplete outcomes.Source: §5.3 · p. 12

5.4 Original failure and descriptive supplement

  • The original chain did not close successfully; authenticated descriptive scoring retained failures but met neither execution nor repeatability targets, and no completed same-population v4 score supports comparison.Source: §5.4 · pp. 12–13

6 Results

In the primary pass, 38 scenes produced correct joint estimates, 57 were withheld and one execution was incomplete; registered targets were not all met. Emitted estimates had small errors, but stress outcomes and repeatability remain bounded by the study design.

6.1 Primary-pass accuracy, withholding and incomplete execution

  • 38 correct estimates, zero wrong estimates, 57 valid withholdings and one incomplete execution yielded 39.1% correct coverage on supported scenes; the registered conjunction failed.Core coverage missed its target by seven successes, combined triaxial coverage by four, and DN coverage by one; two attempts violated the failure-free target.Source: Table 2 · p. 13, Figure 2 · p. 13, §6.1 · pp. 13–14

6.2 Errors among emitted estimates

  • Median sign-only axis error was 0.240° and median rate error was 0.00169% among 38 emissions, but these selected-subset statistics do not describe all scenes.The maximum axis error was 1.997° and maximum rate error was 0.03347%; acceptance prevents assigning the error distribution to withheld scenes or inferring sensor-level calibration.Source: §6.2 · p. 14, Table 3 · p. 14

6.3 Stress outcomes

  • All low-light, biased-intrinsics and optical-axis-misalignment stress scenes were withheld, while some other interventions yielded correct estimates; four scenes per intervention cannot establish robust rejection.Stress correct coverage slightly exceeded supported coverage, but the small, differing strata do not support a matched intervention contrast or significance claim.Source: §6.3 · p. 15, Table 4 · p. 15

6.4 Repetition and the incomplete scene

  • The second pass matched aggregate outcome counts, with scientific equivalence established for 95 pairs and one pair non-comparable because both attempts were incomplete.The incomplete supported-DN scene failed during preparation of block 7 in both attempts; the frozen comparator defines scientific equality, and repetition on one pinned environment is not external reproduction.Source: §6.4 · p. 15

6.5 What zero observed errors can support

  • Under an idealized iid Bernoulli model, zero errors imply illustrative one-sided 95% upper bounds of 7.58% for 38 emissions and 11.29% for 25 supported emissions, not calibrated risk estimates.Fixed strata, repeated geometries, shared rendering assumptions and selection by a complex gate limit interpretation; counting the repeated pass as independent emissions is unjustified.Source: §6.5 · pp. 15–16

7 Historical corrections and negative controls

Historical audits show that permissive axis scoring and motion-dependent acquisition cadence could distort evaluation, while coherent feature bias and sampling aliases can undermine otherwise good fits. Numerical refinements improved objectives locally but did not establish improved physical accuracy.

7.1 An axis metric can erase a physical error

  • 11/13, rather than 13/13, development cases were correct under sign-only scoring; a depth-reflected axis can be orthogonal despite zero error under a reflection quotient.The intended physical quantity, not estimator branch confusion, must determine the scoring equivalence.Source: §7.1 · p. 16, Figure 4 · p. 16

7.2 Acquisition cadence and opened development data

  • 25/48, rather than the historical 38/48, scored correct after axis correction; motion-dependent cadence still exposed a rate proxy, so the final generator fixed frame count independently of motion.The archived-bank replay's 51/96 candidate readouts were development evidence, not a new-study result; historical populations are not additional independent samples.Source: §7.2 · p. 16, Table 5 · p. 17

7.3 Feature identity, closure and aliasing

  • Coherent wrong edges can preserve fit quality, while regular sampling produced a 7.90586°/s alias for a true 25.90586°/s rate at 20 s spacing.The later lattice guard addresses the demonstrated alias but does not establish uniqueness for irregular samples or all trajectories; these historical and analytical probes do not establish flight prevalence.Source: §7.3 · p. 17

7.4 Numerical improvement is not physical improvement

  • 7.85 × 10−9 was the median relative SSE decrease after restarting 36 converged checkpoints, while objective callbacks increased 30.7%.These local numerical effects neither reduced error against truth nor rescued failed fits or added coverage; the evaluated candidate was unchanged.Source: §7.4 · pp. 17–18

8 Discussion

The discussion presents an inspectable synthetic case study with a functioning subset, but its registered coverage and execution requirements remain unmet. It distinguishes numerical, temporal, reproducibility, and physical confidence, identifies validity boundaries, and proposes targeted evaluations to address them.

8.1 What is learned

  • The pipeline produces precise, repeatable joint estimates for some new synthetic scenes, yet fails its minimum-coverage and execution requirements.The authors caution that reporting only accepted precision overstates usefulness, while reporting only the failed target hides the functioning subset.Source: §8.1 · p. 18

8.2 Four distinct notions of confidence

  • Numerical convergence, temporal stability, reproducibility, and physical correctness answer distinct questions; none of the first three alone implies the fourth.A finite surviving-alternative list describes the search performed, not all possible shapes, features, or explanations of the images.Source: §8.2 · p. 18

8.3 Threats to validity

  • Internal validation remains bounded because reference and inference share modeling premises, and whole-video preprocessing makes temporal checks non-independent held-out pixel tests.Prospective scenes protect against direct reuse of final outcomes but do not eliminate shared model misspecification.Source: §8.3 · pp. 18–19
  • The study has only 96 scene draws from four assemblies, no real sensor images, and no completed same-population predecessor or external-estimator comparison.The authors also state that the small conditional-error summaries should not be extrapolated to deployment distributions.Source: §8.3 · p. 19
  • Image-to-report inference was not independently reproduced, and repeating it still requires omitted images, runtime, and remaining dependencies.The package supports checking descriptive arithmetic and recomputing physical errors, but lacks a public archival DOI for the complete evidence at this manuscript version.Source: §8.3 · p. 19

8.4 Experiments that would materially strengthen the claim

  • Proposed strengthening experiments include matched estimator arms, annotated feature correspondences, and separate calibration and untouched test splits.These are proposed experiments, not results or commitments fulfilled by the study.Source: §8.4 · pp. 19–20

9 Conclusion

ORBIT-SCOUT is a selective, inspectable implementation for a restricted monocular precession problem, evaluated on a new synthetic population. The study supports an auditable case report, but not state-of-the-art superiority, calibrated confidence, general identifiability, or flight readiness.

9. Conclusion

  • 38 correct joint estimates, zero wrong emissions, 57 withholding decisions, and one incomplete execution occurred in the primary pass; 95 of 96 pairs repeated scientifically equivalently.Registered coverage and complete-execution requirements were not met.Source: §9 · p. 20
  • The transferable conclusion is to distinguish a fit, an accepted estimate, a withheld answer, and a computation that never established a valid answer.The evidence supports descriptive arithmetic and internal scientific repetition, not calibrated confidence or real-image transfer.Source: §9 · p. 20

Data, code and authoring statement

The source package supports inspection of the implementation and checking of descriptive arithmetic, while complete image-to-report reproduction still depends on materials and runtime not included.

Data, code and authoring statement

  • The package includes per-attempt outcomes, protocol and registration materials, frozen source excerpts, scoring tools, and verification scripts; a separate archive supplies reference files, reports, and receipts.The verifier reproduces physical scoring from originals but does not rerun image inference; rendered images and the complete inference runtime are not included, and no public data DOI is claimed.Source: Data, code and authoring statement · p. 20
  • AI tools assisted development, literature discovery, evidence inspection, and drafting; the sole named author retains scientific responsibility and submission decisions.Automated adversarial reviews were software-agent reviews, not human peer review.Source: Data, code and authoring statement · p. 20

A Endpoint correction and its limitations

The endpoint objective reduces correction to a candidate line through a vanishing point, with a separate common-ray constraint. Its geometric fit is not by itself evidence that endpoints share a physical generator or that the loss is an appropriate noise model.

A Endpoint correction and its limitations

B Frozen numerical search envelope

B.1 Cone bank

The search banks use fixed solver envelopes and feed an explicit gate policy. Conditional stability and rival checks qualify the emitted estimates, but they are not calibrated confidence guarantees or tests over independent frames.

B.1 Cone bank

  • The cone bank uses 480 starts—240 uniformly spaced initial frequencies and two depth lifts—with frequency scaled by 360/∆TTRAIN; numerical termination accepts solver success or nonnegative canonical loss at most 10^-20.This floor rule is not a stationarity or accuracy certificate.Source: §B.1 · pp. 21–22

B.2 Triaxial bank

  • The triaxial bank selects 128 starts from diverse cone parents and uses staged least-squares budgets, with an optional refinement that does not guarantee convergence.The scale bound on r is not a 30 degrees/s bound on every reported mean precession.Source: §B.2 · p. 22

C Gate inventory and decision semantics

C.1 Conditional dispersion and rival contrasts

  • Conditional dispersion gates require Rh < 10 degrees and RΩ < 15%, but overlapping training sets and shared features correlate folds, so αnom = 0.01 is not calibrated 99% coverage.Source: §C.1 · pp. 23–24
  • Every retained TRAIN-compatible rival must meet the declared training and held-out contrast conditions; absent rivals pass vacuously, and missing held-out evidence makes comparison incomplete.Source: §C.1 · p. 24
  • Duplicating every frame can change the family-selection score difference without adding independent information, so its dimension penalty is heuristic.Source: §C.1 · p. 24

D Evidence identity and reproduction boundaries

The evidence package distinguishes descriptive rebuilding, physical-metric rescoring, and image-to-report reproduction. Its checks support identity and scoring reproducibility within the stated boundaries, not correctness of the experiment or unrestricted external reproduction.

E Claim boundaries for reading and reuse

D Evidence identity and reproduction boundaries

  • Descriptive arithmetic rebuilds counts and plots without inference, while the physical-metric level rescored archived reference targets and reports.Source: D Evidence identity and reproduction… · p. 25
  • All 76 emitted-attempt error pairs reproduced archived floating-point values exactly, and all 192 outcome labels agreed in the checked environment.This post-closure scoring check was not a new accuracy population or image-inference run.Source: D Evidence identity and reproduction… · p. 25
  • Image-to-report reproduction requires the pinned runtime, frozen candidate, inputs, and evaluation artifacts; the compact package alone cannot perform it.New execution must retain attempted cases and report failures rather than select successful runs.Source: D Evidence identity and reproduction… · p. 25

Related papers

Loading 2610.34819v1…