Pith. sign in

REVIEW 5 major objections 4 minor 50 references

Rolling Shutter Camera Self-Calibration

T0 review · 5 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper presents the first self-calibration method for rolling-shutter cameras, jointly estimating camera intrinsics and the readout time ratio from image sequences without calibration targets.

desk verdict RSSC has a genuinely useful dual-projection formulation, but the 'first self-calibration' claim needs a sharper prior-art argument and the supplementary Jacobians contain an SE(3) error that has to be resolved before the results can be trusted. read the letter →

arxiv 2608.01509 v1 pith:5AOKH4PY submitted 2026-08-02 cs.CV

Yongcong Zhang , Navid Rabbani , Bangyan Liao , Chengbo Wang , Yizhen Lao , Adrien Bartoli This is my paper
classification cs.CV
keywords rollingshutterself-calibrationbundleadjustmentreadouttimeratiocumulativeB-splinescontinuous-timetrajectorycorrectionfieldestimationcameraintrinsics
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

Rolling-shutter cameras expose different image rows at slightly different times, so motion warps images in a way that depends on the camera intrinsics and on the readout time ratio. This paper claims the first method that recovers both quantities from ordinary video, without printed targets, LED panels, or inertial sensors. The method is a self-calibrating bundle adjustment that fuses two complementary views of the rolling-shutter effect: a continuous-time camera trajectory in which every row has its own pose, and a correction-field view that rectifies the image to a global-shutter frame. The fused dual-projection model constrains each 3D point at two timestamps along one shared trajectory, which the authors show improves accuracy and stability across synthetic and real sequences. If the claim holds, any sufficiently textured video from a rolling-shutter camera can serve as its own calibration rig.

What carries the argument

The load-bearing object is the dual-projection model over a shared cumulative cubic B-spline trajectory on $SE(3)$. The trajectory supplies a continuous pose $T(t)$, so the pose of a row at time $t(\gamma, v) = (i + \gamma v/H)T_f$ can be evaluated alongside the pose at a reference global-shutter time $t_i$. For each 3D point, the residual combines the row-dependent RS reprojection at $t(\gamma, v)$ with the rectified GS reprojection at $t_i$, where the rectified measurement comes from Lagrange or Hermite interpolation of the point's 2D trajectory across neighboring frames. That coupling forces $K$ and $\gamma$ to be consistent with both temporal interpretations of the same observation.

What would settle it

Run RSSC-DP on a synthetic video with known $K$ and $\gamma$ but a motion known to be critical for self-calibration (e.g., pure rotation about the camera center, or a planar scene with no parallax) while using the paper's standard initialization; if the recovered $\gamma$ stays at the initial value or the intrinsics drift, the general-video claim fails. A cheap screening test is to sweep $\gamma$ from near 0 to near 1 at fixed trajectory and noise level and record the recovery rate.

Watch

Extended reading notes

Core claim

The central claim is that rolling-shutter self-calibration is possible: a bundle adjustment that optimizes the intrinsics $K$, the readout time ratio $\gamma$, the control poses of a cumulative B-spline trajectory, and the 3D points, with each observation projected twice — once at the row-dependent timestamp and once at a reference global-shutter timestamp — recovers $K$ and $\gamma$ without markers or hardware. The component models are RSSC-TE (continuous-time trajectory estimation) and RSSC-CE (correction-field estimation with quadratic or Hermite interpolation), and their union RSSC-DP ties the two projections to the same trajectory. In real-data experiments the proposed variants consiste

Load-bearing premise

The method recovers $K$ and $\gamma$ from 'a general video' only if the motion is non-degenerate and the initial global-shutter reconstruction, started with $\gamma=0.5$, lies in the basin of convergence; the paper defers the degeneracy analysis that would say which motions satisfy this.

Editorial extensions

If this is right

  • Because the method is target-free, any sufficiently textured, non-degenerate video from a rolling-shutter camera could replace checkerboards, LED panels, and inertial sensors as the calibration input.
  • The recovered intrinsics and readout ratio can initialize structure-from-motion, odometry, and image rectification with correct rolling-shutter parameters instead of a global-shutter assumption.
  • The ablation shows a practical trade-off: correction-field variants converge fastest, trajectory estimation is stable under smooth motion, and the dual-projection variant is the most accurate across conditions.
  • The supplementary comparison with target-based rolling-shutter calibration indicates the target-free estimates are competitive, so the method could also be used in controlled settings where a pattern is inconvenient.

Reading between the lines

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

  • The authors explicitly defer degeneracy analysis, so 'general video' should be read as 'non-degenerate motion'; a catalogue of motions that make the readout ratio unobservable is the immediate next test.
  • Because the pipeline starts from a global-shutter initializer, it likely inherits that initializer's failure modes under strong rolling-shutter distortion; an alternating initialization-and-refinement loop is a testable extension.
  • The dual-timestamp projection idea is not tied to B-splines specifically; it transfers to any continuous-time motion model and, in principle, to sensors with per-row or per-line temporal sampling.
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.

Referee Report

5 major / 4 minor

Summary. The paper presents RSSC, a framework for rolling-shutter (RS) camera self-calibration from image sequences without calibration targets. It combines two complementary models: RSSC-TE, which represents the camera trajectory with cumulative cubic B-splines and evaluates row-dependent poses; and RSSC-CE, which rectifies RS observations to global-shutter (GS) timestamps via 2D interpolation. These are unified in RSSC-DP, a weighted bundle-adjustment objective that constrains each 3D point by both the original RS projection and the rectified GS projection. The method jointly estimates intrinsics K, readout time ratio γ, poses, and 3D points. Experiments on synthetic data and the WHU-RSVI/TUM-RSVI datasets report improved intrinsic and trajectory accuracy over GS baselines, and the supplementary compares with a target-based RS calibration method.

Significance. If the technical issues below are resolved, the paper would be a useful contribution: it is a systematic attempt at target-free RS self-calibration in a BA framework, the dual-projection idea is elegant, and the open-source project page plus complete supplementary Jacobians are good practice. The paper also provides falsifiable predictions in the form of numeric error tables on standard datasets. However, the central claims currently outrun the evidence: the analytic Jacobians are suspect and internally inconsistent, degeneracy analysis is explicitly deferred, and the real-data evaluation lacks uncertainty quantification and a true RS self-calibration baseline. With a corrected Jacobian derivation and a sharper evaluation, the work could be publishable.

major comments (5)
  1. [Supp. Sec. 2.1, Eq. (16)] The Jacobian derivation is internally inconsistent and, as written, likely incorrect. Supp. Eq. (16) states ∂a_k/∂ξ_{i+k-1} = -∂a_k/∂ξ_{i+k-2} for k=1,2,3, where a_k = B_k(s) log(T_{i+k-2}^{-1}T_{i+k-1}). The derivative with respect to the two endpoints involves different Lie derivatives: a right perturbation of T_{i+k-1} gives B_k J_r^{-1}(Ω), while a perturbation of T_{i+k-2} introduces an additional adjoint factor, so the two derivatives are not negatives except in the zero-relative-motion limit. Indeed, Eq. (17) uses J_r^{-1}(Ω_{i+1}) Ad_{T_i^{-1}} for the ξ_i block, which contradicts Eq. (16). Since Algorithm 1 (step 28) forms J^T J from these blocks, the correctness of the reported convergence and accuracy results depends on this discrepancy. Please provide a corrected derivation or verify all Jacobians against numerical/automatic differentiation.
  2. [Sec. 6 (and Sec. 5.1)] The paper claims target-free self-calibration from 'general video', but no degeneracy analysis is provided; Sec. 6 explicitly defers it to future work. The synthetic study varies trajectory smoothness, γ, and measurement noise, but does not vary motion type (e.g., pure rotation, constant velocity, planar motions) that are known to be critical for RS SfM and self-calibration (see Refs. [4,16]). A concrete test would be to perturb the trajectory to such critical sequences and report γ/K observability or the basin of convergence around the COLMAP initialization used in Sec. 5.2. Without this, the central claim is formally unsupported.
  3. [Sec. 5.2, Table 3 and Supp. Tables 3-4] The real-data claims are stronger than the evidence. Table 3 reports only medians over sequences, with no spread, confidence intervals, or repeated-run variation; the supplementary per-sequence tables show large variance (e.g., TUM-RSVI6 RSSC-CEQ has ATE 0.1348 while RSSC-TE is 0.0085, and RSSC-CEH is 0.3363). The main baselines are all global-shutter methods; the only RS-aware method, RSCC, appears only in the supplementary Table 2, requires a calibration pattern, and is evaluated on only two short sequences, where the proposed methods are generally worse on γ. Thus 'consistently outperform' and 'competitive with target-based RS calibration' are not established at this level of evidence.
  4. [Sec. 4.2, Eqs. (8)-(15)] The CE residual is not an independent measurement. The rectified observation GSm_j^i(t_i) is computed from the observed points and timestamps t(γ, v) via Eq. (8) or (10); the interpolation weights depend on γ (Eq. (9)/(12)-(13)). Minimizing the residual (15) can therefore shrink the interpolation error by moving γ rather than by improving geometric consistency. RSSC-DP inherits this effect through its CE term (Eq. (18)). The synthetic results show γ errors for RSSC-CE growing as γ decreases, which is consistent with this coupling. The paper should analyze this bias or justify that the BA does not exploit it.
  5. [Sec. 2 (Related Work) and References] The claimed novelty 'first self-calibration method for RS cameras' is not properly positioned. Reference [16] (Ito & Okatani, CVPR 2017) is cited in the bibliography but never discussed in Sec. 2. Its title explicitly concerns 'self-calibration-based approach to critical motion sequences of rolling-shutter structure from motion'. If that method estimates RS intrinsics/readout parameters from target-free sequences, then the 'first' claim in the Abstract and Conclusion is inaccurate; if not, the distinction should be stated explicitly. Please add a comparative discussion.
minor comments (4)
  1. [Fig. 6 and Table 3] Fig. 6 uses labels RSSC-ET and RSSC-ECQ, while the text/table uses RSSC-TE and RSSC-CEQ; please unify. Also, the caption should identify which RSSC variants are shown.
  2. [Algorithm 1, steps 20-24] Step 20 forms residual e = [TE; λ·CE], but it is unclear whether λ scales the residual or the normal equations; the Jacobian assembly should be specified consistently.
  3. [Supp. Table 2] The title has a typo ('Comparison with RS-aware approache'). Please also define the abbreviations in the table caption.
  4. [Sec. 3.1, Eq. (1)] The definition of γ = T_readout/T_f is given, but Eq. (1) does not make explicit whether exposure time is included in T_readout; please clarify the relationship between row timestamp and exposure.

Circularity Check

1 steps flagged · score 2.0 of 10

No load-bearing circularity; one mild self-reference in the RSSC-CE rectified-observation residual, whose target is constructed from the readout ratio being estimated.

  1. self definitional [Sec. 4.2, Eqs. (8), (9), (15); used in RSSC-CE and RSSC-DP]
    "GSm_j^i(t_i) = L_0(t_i)m_{i-1}^j + L_1(t_i)m_i^j + L_2(t_i)m_{i+1}^j, ... where the Lagrange basis functions are: L0(ti)=... (with t(γ,v) from Eq. (1)). CEe_j^i = GSm_j^i(t_i) − π(K GST_i P^j)."

    The rectified observation GSm used as the target in the CE residual is not a raw measurement; it is constructed from the raw points via interpolation weights that depend on the readout time ratio γ (through t(γ,v) in Eq. (9) and Eq. (1)). Thus γ appears on both sides of the residual: it determines the 'measured' corrected point and is then optimized to match the projection. This makes the CE term a self-consistency constraint rather than an independent observation of γ. The effect is partial—RSSC-TE and the TE term in RSSC-DP use raw observations and provide independent anchoring, and real-data comparisons against ground truth validate the result externally—but the CE residual is not an independent measurement by construction.

full rationale

The core self-calibration claim is not circular: RSSC-TE minimizes a reprojection error between raw RS observations and a continuous-time B-spline trajectory, with γ entering only through the row timestamp in the pose interpolation (Eq. (7)); this is a standard geometric fit, not a reduction to the target. The real-data experiments are benchmarked against external datasets (WHU-RSVI, TUM-RSVI) with ground-truth intrinsics and trajectories, providing independent anchoring. The self-citations to prior work (e.g., [31] for quadratic interpolation, [24] for RS BA) are technique citations, not load-bearing uniqueness theorems. The mild self-reference flagged above in RSSC-CE/-DP is real but does not sink the central claim, because the TE component and the raw observations remain independent. The paper's explicit deferral of degeneracy analysis (Sec. 6, 'Future work includes ... studying degeneracies in RS camera self-calibration') and the potential Jacobian issue in Supp. Eq. (16) are correctness/completeness concerns, not circularity. Overall, the derivation chain is self-contained against external benchmarks; score 2 reflects the one constructed-measurement dependency rather than any forced equivalence.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method rests on a row-linear RS model, B-spline smoothness, local 2D trajectory interpolation, and the unproven assumption that a GS COLMAP initialization plus non-degenerate motion lead to convergence. The only hand-set free parameter in the method itself is the DP weight lambda; gamma initialization is a practical choice. No new physical entities are introduced.

free parameters (2)
  • lambda (DP weight) = 0.5
    Eq. (19): hand-set weight balancing TE and CE residuals; no sensitivity analysis is reported.
  • gamma initialization = 0.5
    Sec. 5.2: readout time ratio initialized to 0.5 in all real data experiments; convergence is not analyzed.
assumptions (4)
  • domain assumption Row-linear readout model with a single readout ratio gamma (Eq. 1)
    Assumes uniform row delay; non-uniform or per-row readout is excluded.
  • domain assumption Cumulative cubic B-spline trajectory smoothness (Eq. 3)
    Poses are interpolated from four control poses; jerky or very fast motion violates this assumption.
  • domain assumption Quadratic or cubic interpolation of 2D point trajectories (Eqs. 8, 10)
    The rectified target depends on locally polynomial image motion; this is a modeling prior, not derived from 3D geometry.
  • ad hoc to paper Non-degenerate motion and basin of convergence around the COLMAP initialization (Sec. 5.2)
    No degeneracy analysis is given; Sec. 6 defers studying degeneracies in RS camera self-calibration, so the 'general video' claim is assumed rather than shown.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Rolling Shutter Camera Self-Calibration." pith.science (2026). https://pith.science/paper/5AOKH4PY

@misc{pith2026260801509,
  author       = {Pith},
  title        = {Pith review of: Rolling Shutter Camera Self-Calibration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5AOKH4PY}},
  note         = {Machine review of arXiv:2608.01509}
}
read the original abstract

Rolling shutter (RS) cameras are widely used in consumer devices, but their row-wise exposure causes distortions under motion, making geometric 3D vision problems dependent on both camera intrinsics and readout time ratio. Existing RS calibration methods rely on calibration targets or specialised hardware, limiting their use in unconstrained settings. We present the first self-calibration method for RS cameras that directly estimates camera intrinsics and the readout time ratio from image sequences, without requiring calibration targets. The method is implemented as a self-calibrating bundle adjustment (BA), which critically depends on the RS imaging model. We combine two known complementary models. The first formulates RS imaging as continuous-time trajectory estimation under a row-wise pose representation. The second interprets RS images as temporally distorted global shutter (GS) images and requires to estimate correction fields. The combination is non-trivial and results in a unified dual-projection model, in which each 3D point is simultaneously constrained at both row-dependent and reference timestamps along a shared continuous trajectory, enforcing stronger geometric and temporal consistency. Extensive simulations analyse the applicability of several implementations under varying conditions, and real data experiments demonstrate the accuracy, robustness, and practical effectiveness of the proposed approach.

Figures

Figures reproduced from arXiv: 2608.01509 by the authors.

Figure 1
Figure 1. Computation pipeline of the proposed Rolling Shutter Camera Self-Calibration (RSSC). Given a sequence of RS images, an initial reconstruction of the camera in￾trinsics K, camera poses T, and 3D point coordinates P is first obtained. It is then refined by the RSSC framework, where a self-calibrating BA jointly estimates the read￾out time ratio γ. The estimated RS camera parameters can subsequently be used for downstr… view at source ↗
Figure 2
Figure 2. (a) Two perspectives for modeling RS images. Left: the RS image is modeled as a composition of multiple independently captured single-row GS images; Right: the RS image is modeled as a single GS image with distortion. (b) The proposed dual￾projection model projects each 3D point twice, onto its original RS 2D point and onto the rectified GS 2D point. B-spline representation, assigning a unique camera pose to each fe… view at source ↗
Figure 3
Figure 3. Illustration of the RS imaging and pixel-wise exposure modeling. (a) Exposure mechanism between two RS consecutive frames. (b) Pixel exposure time formulation within an RS image. 3.1 Rolling Shutter Camera Projection Models Unlike a GS camera, an RS camera captures images row by row. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Two different fitting strategies: (a) Representing the camera trajectory as a function of γ. (b) Representing the 2D points as a function of γ. Each point along the trajectory is influenced by four consecutive control points. Specifically, for a given time t ∈ [ti , ti…
Figure 5
Figure 5. Figure 5: Intrinsic parameters and readout time ratio errors for the proposed RS camera self-calibration methods under varying experimental conditions, with, from left to right, trajectory smoothness, readout time ratio, and measurement noise. 4.4 Discussion The two constraints …
Figure 6
Figure 6. Figure 6: Trajectory comparison on the WHU-RSVI [6] and TUM-RSVI [8] datasets. Each row denotes a different trajectory sequence, while each column corresponds to a specific method. Selected local regions (highlighted by dashed boxes) are magnified and displayed within solid boxe…
Figure 1
Figure 1. Figure 1: Trajectory comparison on the WHU-RSVI [1] and TUM-RSVI [2] datasets. Each row denotes a different trajectory sequence, while each column corresponds to a specific method. Selected local regions (highlighted by dashed boxes) are magnified and displayed within solid boxe…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 45 canonical work pages

  1. [16]

    International journal of computer vision , volume=

    Shape and motion from image streams under orthography: a factorization method , author=. International journal of computer vision , volume=. 1992 , publisher=

    work page 1992

  2. [1]

    , author=

    Spline Fusion: A continuous-time representation for visual-inertial fusion with application to rolling shutter cameras. , author=. BMVC , volume=

    work page

  3. [2]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Towards nonlinear-motion-aware and occlusion-robust rolling shutter correction , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    work page

  4. [3]

    Astronomical and space optical systems , volume=

    Fundamental performance differences between CMOS and CCD imagers: part III , author=. Astronomical and space optical systems , volume=. 2009 , organization=

    work page 2009

  5. [4]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Fast rolling shutter correction in the wild , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2023 , publisher=

    work page 2023

  6. [5]

    arXiv preprint cs/0503076 , year=

    Geometric models of rolling-shutter cameras , author=. arXiv preprint cs/0503076 , year=

    arXiv

  7. [6]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Rolling shutter camera absolute pose , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2019 , publisher=

    work page 2019

  8. [7]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Rolling shutter camera relative pose: Generalized epipolar geometry , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    work page

Show all 50 references
  1. [8]

    European Conference on Computer Vision , pages=

    Degeneracies in rolling shutter sfm , author=. European Conference on Computer Vision , pages=. 2016 , organization=

    2016

  2. [9]

    2012 IEEE Conference on Computer Vision and Pattern Recognition , pages=

    Rolling shutter bundle adjustment , author=. 2012 IEEE Conference on Computer Vision and Pattern Recognition , pages=. 2012 , organization=

    2012

  3. [10]

    Pattern Recognition Letters , volume=

    Robustified structure from motion with rolling-shutter camera using straightness constraint , author=. Pattern Recognition Letters , volume=. 2018 , publisher=

    2018

  4. [11]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Revisiting rolling shutter bundle adjustment: Toward accurate and fast solution , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  5. [12]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages=

    Rolling shutter camera calibration , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages=

  6. [13]

    IEEE Sensors Journal , volume=

    Continuous-time spatiotemporal calibration of a rolling shutter camera-IMU system , author=. IEEE Sensors Journal , volume=. 2022 , publisher=

    2022

  7. [14]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages=

    Self-calibration-based approach to critical motion sequences of rolling-shutter structure from motion , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages=

  8. [15]

    2023 , publisher=

    Automated rolling shutter calibration with an LED panel , journal=. 2023 , publisher=

    2023

  9. [17]

    2003 , publisher=

    Multiple view geometry in computer vision , author=. 2003 , publisher=

    2003

  10. [18]

    Proceedings of the IEEE International Conference on Computer Vision , pages=

    Rolling-shutter-aware differential sfm and image rectification , author=. Proceedings of the IEEE International Conference on Computer Vision , pages=

  11. [19]

    IEEE Transactions on Intelligent Vehicles , year=

    Ego-motion Estimation for Vehicles with a Rolling Shutter Camera , author=. IEEE Transactions on Intelligent Vehicles , year=

  12. [20]

    IEEE transactions on image processing , volume=

    Analysis and compensation of rolling shutter effect , author=. IEEE transactions on image processing , volume=. 2008 , publisher=

    2008

  13. [21]

    Sensors , volume=

    Calibration and noise identification of a rolling shutter camera and a low-cost inertial measurement unit , author=. Sensors , volume=. 2018 , publisher=

    2018

  14. [22]

    European conference on computer vision , pages=

    Autocalibration from planar scenes , author=. European conference on computer vision , pages=. 1998 , organization=

    1998

  15. [23]

    International Journal of computer vision , volume=

    Self-calibration of a moving camera from point correspondences and fundamental matrices , author=. International Journal of computer vision , volume=. 1997 , publisher=

    1997

  16. [24]

    2015 IEEE International Conference on Image Processing (ICIP) , pages=

    Deepfocal: A method for direct focal length estimation , author=. 2015 IEEE International Conference on Image Processing (ICIP) , pages=. 2015 , organization=

    2015

  17. [25]

    Proceedings of the 15th ACM SIGGRAPH European Conference on Visual Media Production , pages=

    DeepCalib: A deep learning approach for automatic intrinsic calibration of wide field-of-view cameras , author=. Proceedings of the 15th ACM SIGGRAPH European Conference on Visual Media Production , pages=

  18. [26]

    2022 International Conference on Robotics and Automation (ICRA) , pages=

    Self-supervised camera self-calibration from video , author=. 2022 International Conference on Robotics and Automation (ICRA) , pages=. 2022 , organization=

    2022

  19. [27]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Deep geometry-aware camera self-calibration from video , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  20. [28]

    European conference on computer vision , pages=

    Camera self-calibration: Theory and experiments , author=. European conference on computer vision , pages=. 1992 , organization=

    1992

  21. [29]

    International workshop on vision algorithms , pages=

    Bundle adjustment—a modern synthesis , author=. International workshop on vision algorithms , pages=. 1999 , organization=

    1999

  22. [30]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Structure-from-motion revisited , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  23. [31]

    European Conference on Computer Vision , pages=

    Geocalib: Learning single-image calibration with geometric optimization , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  24. [32]

    Structure-from-Motion Revisited , booktitle=

    Sch\". Structure-from-Motion Revisited , booktitle=

  25. [33]

    IEEE Access , volume=

    The WHU rolling shutter visual-inertial dataset , author=. IEEE Access , volume=. 2020 , publisher=

    2020

  26. [34]

    IROS , year=

    Rolling-shutter modelling for direct visual-inertial odometry , author=. IROS , year=

  27. [35]

    International journal of computer vision , volume=

    A theory of self-calibration of a moving camera , author=. International journal of computer vision , volume=. 1992 , publisher=

    1992

  28. [36]

    Proceedings of the IEEE Conference on Advanced Video and Signal Based Surveillance, 2003

    A survey of camera self-calibration , author=. Proceedings of the IEEE Conference on Advanced Video and Signal Based Surveillance, 2003. , pages=. 2003 , organization=

    2003

  29. [37]

    Proceedings of the European conference on computer vision (ECCV) , pages=

    Fisheyerecnet: A multi-context collaborative deep network for fisheye image rectification , author=. Proceedings of the European conference on computer vision (ECCV) , pages=

  30. [38]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Learning to calibrate straight lines for fisheye image rectification , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  31. [39]

    Proceedings of the IEEE/CVF International conference on computer vision , pages=

    Ctrl-c: Camera calibration transformer with line-classification , author=. Proceedings of the IEEE/CVF International conference on computer vision , pages=

  32. [40]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Deep single image camera calibration with radial distortion , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  33. [41]

    2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=

    Degeneracy in self-calibration revisited and a deep learning solution for uncalibrated slam , author=. 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=. 2019 , organization=

    2019

  34. [42]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Depth from videos in the wild: Unsupervised monocular depth learning from unknown cameras , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  35. [43]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  36. [44]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Self-calibrating neural radiance fields , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  37. [45]

    arXiv preprint arXiv:1806.04807 , year=

    Ba-net: Dense bundle adjustment network , author=. arXiv preprint arXiv:1806.04807 , year=

    arXiv

  38. [46]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Vggsfm: Visual geometry grounded deep structure from motion , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  39. [47]

    2025 International Conference on 3D Vision (3DV) , pages=

    Flowmap: High-quality camera poses, intrinsics, and depth via gradient descent , author=. 2025 International Conference on 3D Vision (3DV) , pages=. 2025 , organization=

    2025

  40. [48]

    Proceedings of the European Conference on Computer Vision (ECCV) , pages=

    Direct sparse odometry with rolling shutter , author=. Proceedings of the European Conference on Computer Vision (ECCV) , pages=

  41. [49]

    2016 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Direct semi-dense SLAM for rolling shutter cameras , author=. 2016 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2016 , organization=

    2016

  42. [50]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    From two rolling shutters to one global shutter , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.