Pith. sign in

REVIEW 2 major objections 8 minor 32 references

Synthetic video cuts focal length error by 24%

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · glm-5.2

2026-07-07 12:37 UTC pith:RWI7LIQK

load-bearing objection Solid dataset contribution with a thin-lens synthetic pipeline and a practical calibration improvement; the finetuning claim is overstated and needs rewording plus ideally multiple seeds. the 2 major comments →

arxiv 2607.05389 v1 pith:RWI7LIQK submitted 2026-07-06 cs.CV

InFlux++: Real and Synthetic Data for Estimating Dynamic Camera Intrinsics

Erich Liang , Caleb Kha-Uong , Chinmaya Saran , Sreemanti Dey , David W. Liu , Junhan Ouyang , Benjamin Zhou , Jia Deng This is my paper
classification cs.CV
keywords dynamic camera intrinsicssynthetic datacamera calibrationlens distortiondefocus blurthin lens modelprocedural generationbenchmark
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper presents InFlux++, a two-part data suite for predicting per-frame camera intrinsics from RGB video. The suite includes InFlux++ Synth, a procedurally generated synthetic video dataset of 441K+ frames across 1841 videos with perfectly accurate per-frame ground truth intrinsics, and InFlux++ Real, a real-world benchmark of 514K+ frames across 334 videos that extends the prior InFlux benchmark with broader scene and camera motion diversity. The synthetic dataset uses the thin lens model to render physically realistic defocus blur and lens breathing, and applies Brown-Conrady lens distortion, providing optical effects absent from prior synthetic datasets. The central claim is that finetuning an existing state-of-the-art intrinsics prediction method (AnyCalib) on InFlux++ Synth consistently improves focal length estimation on both InFlux++ Real and InFlux, raising combined fx recall@10% from 25.2% to 31.2% and fy recall@10% from 25.2% to 31.3%, suggesting synthetic supervision is a viable path for RGB-based dynamic intrinsics prediction.

Core claim

Finetuning AnyCalib on the InFlux++ Synth synthetic dataset improves focal length estimation across both the InFlux++ Real and InFlux benchmarks. The combined fx recall@10% rises from 25.2% to 31.2% and fy recall@10% from 25.2% to 31.3%. However, principal point (cx, cy) and endpoint error (EPE) metrics worsen after finetuning, indicating that synthetic supervision helps with focal length but does not uniformly improve all intrinsics parameters.

What carries the argument

The thin lens model (1/LFL = 1/CFL + 1/LTO) is used to parameterize camera focal length as a function of lens focal length and lens-to-object distance, producing realistic lens breathing and defocus blur in Blender's Cycles renderer. LFL and LTO are varied over time via bounded random walks with Bézier interpolation. Brown-Conrady lens distortion is applied via a data loader. For the real benchmark, a new board-based calibration procedure replaces drone-based calibration for large field-of-view spatial footprints.

Load-bearing premise

The thin lens model used in Blender's Cycles renderer accurately captures the relationship between lens focal length, camera focal length, and lens-to-object distance, and the resulting defocus blur and lens breathing provide a supervisory signal that transfers to real-world footage. If real lens behavior deviates substantially from this model (as with compound zoom lens optics or chromatic aberrations), the synthetic training signal may not transfer.

What would settle it

If finetuning on InFlux++ Synth produced no improvement, or worsened focal length estimation on the real benchmarks, the core claim that synthetic supervision is promising for dynamic intrinsics prediction would be undermined.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Synthetic data with physically realistic optical effects (defocus blur, lens breathing, lens distortion) can serve as scalable supervision for camera calibration models, reducing reliance on expensive real-world capture with specialized lens hardware.
  • If the thin lens model adequately approximates real zoom lens behavior, then synthetic-to-real transfer for intrinsics prediction is practical without domain randomization on the optical model itself.
  • The improvement in focal length but degradation in principal point and EPE metrics suggests that training objectives and loss functions may need redesign to jointly optimize all intrinsics parameters, particularly distortion near image boundaries.
  • The board-based large-FSF calibration procedure simplifies benchmark construction, potentially lowering the barrier for new lens calibration and enabling broader lens diversity in future datasets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The mismatch between AnyCalib's spherical ray loss and the EPE metric's planar projection may be a general problem for any method optimizing angular ray losses; switching to a loss that weights boundary pixels more heavily could simultaneously improve distortion and EPE without additional data.
  • If compound zoom lens optics deviate substantially from the thin lens model, a more complex optical model (e.g., thick lens or multi-element lens simulation) might be needed to close the remaining sim-to-real gap, particularly for distortion prediction.
  • The fact that finetuning improves focal length but worsens principal point estimation on InFlux++ Real (which has more off-center principal points) suggests the synthetic data's principal point distribution may not match the real benchmark's, and targeted augmentation of principal point offsets could help.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 8 minor

Summary. This paper presents InFlux++, a data suite for dynamic camera intrinsics prediction consisting of two components: (1) InFlux++ Synth, a large-scale procedurally generated synthetic video dataset with 441K+ frames across 1841 videos, featuring per-frame ground truth intrinsics with diverse zoom, focus, lens distortion, and defocus blur; and (2) InFlux++ Real, a real-world benchmark extending InFlux with 514K+ frames across 334 videos with broader scene and camera motion diversity. The synthetic dataset is built on Infinigen and uses the thin lens model to parameterize camera focal length via lens focal length and lens-to-object distance, producing realistic lens breathing effects. The real benchmark uses a simplified board-based calibration procedure for large field-of-view spatial footprints, replacing the drone-based approach from InFlux. The authors evaluate seven baseline methods and finetune AnyCalib on InFlux++ Synth, reporting improved focal length recall but degraded principal point and EPE metrics.

Significance. The paper addresses a genuine gap in dynamic camera intrinsics research: the scarcity of training data with per-frame intrinsics variation and the limited diversity of existing benchmarks. The dataset construction methodology is well-documented, including the thin lens parameterization (Sec. 4.1.3), bounded random walks for LFL/LTO variation (Algorithm 1), the corrected EPE visibility filter (Sec. 5.1, Appendix K.1), and the LUT-reliable EPE recall formulation (Appendix K.2). The large-scale board-based calibration procedure for large FSF settings (Sec. 4.2.1) is a practical contribution that simplifies data collection. The public release of data, code, a data loader, and a live evaluation server is commendable and strengthens the work's reproducibility. The finetuning experiment provides a falsifiable, if preliminary, signal that synthetic supervision transfers to real-world focal length estimation.

major comments (2)
  1. Sec. 5.3 and Tab. 3: The checkpoint is selected based on 'the lowest CFL percent error on the validation split of InFlux++ Real.' This selection criterion directly targets the CFL metric family (fx/fy recall) that is then reported as improved, while EPE recall@50px drops from 25.1% to 18.7% and cx recall@1% drops from 76.5% to 67.2% (Tab. 3, Combined). The claim that finetuning 'consistently improves focal length estimation' is over-scoped: the improvement is consistent only for the metric the checkpoint was selected to optimize, and the simultaneous degradation across all other intrinsics metrics is consistent with a trade-off induced by the selection criterion. Without multiple random seeds or variance estimates, it is unclear whether the +6% fx recall@10% gain is robust or could arise from checkpoint selection variance in a single run. The authors should either (a) report results from
  2. Sec. 5.3, Tab. 3: The abstract states that finetuning 'consistently improves focal length estimation across both InFlux++ Real and InFlux.' However, Tab. 3 shows that on InFlux, EPE recall@50px drops from 34.1% to 27.9% and EPE recall@10px drops from 2.95% to 2.09%. While fx/fy recall does improve, the simultaneous and substantial degradation in EPE—a metric that directly measures the geometric quality of predicted intrinsics—undermines the framing of the result as a straightforward improvement. The paper should more carefully qualify the claim in the abstract and Sec. 5.3 to acknowledge that the improvement is specific to CFL metrics and comes at the cost of EPE and principal point accuracy, rather than presenting it as a general improvement in intrinsics estimation.
minor comments (8)
  1. Sec. 4.1.3: The thin lens equation is written as '1/LFL = 1/CFL + 1/LTO' but is displayed in a format that could be misread. Standard convention writes 1/f = 1/u + 1/v where f is the lens focal length. Consider clarifying the correspondence between the standard thin lens variables and the paper's LFL/CFL/LTO terminology to avoid confusion for readers unfamiliar with the specific parameterization.
  2. Tab. 2: The table is dense and the formatting of very small numbers (e.g., '5.68e-1', '8.53e-1') is inconsistent with the percentage formatting of other entries. Consider using a uniform format (e.g., '0.568' or '0.57%') for readability.
  3. Fig. 4: The k1 and k2 distributions are described as 'heavily centered around 0.' It would help to state the sampling parameters (means, standard deviations) explicitly in the caption or text, rather than only in Appendix F, so readers can assess the distortion diversity without consulting the supplement.
  4. Sec. 4.1.4: The LFL bounds of [8mm, 100mm] are stated without justification. A brief note on why this range was chosen (e.g., to match the range of real lenses in InFlux++ Real) would help readers understand the design decision.
  5. Tab. 1: The scene category counts for InFlux total 386 videos, while the text states InFlux++ Real has 334 videos. The InFlux total differs from the 386 stated in the original InFlux paper's typical citation. Consider verifying this count or noting if some InFlux videos were excluded.
  6. Appendix L.2: The learning rate schedule description mentions 'MultiStepLR decay with factor gamma=0.3 at epochs 4 and 8' but does not specify the final learning rate or whether any learning rate warmup restart was used. Minor, but completeness would help reproducibility.
  7. Sec. 5.1: The LUT-reliable EPE recall threshold tau=0.95 is introduced without discussion of sensitivity. A brief note on how the choice of tau affects the set of reliable frames would strengthen the metric's justification.
  8. The paper uses 'consistently improves' in the abstract, contributions list (Sec. 1), and Sec. 5.3. Given the mixed results in Tab. 3, consider softening to 'improves focal length estimation' without 'consistently,' or explicitly scoping the claim to CFL metrics.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive feedback. Both major comments concern the framing and robustness of our finetuning experiment (Sec. 5.3, Tab. 3). We agree that the current presentation overstates the generality of the improvement and that the checkpoint selection criterion introduces a bias toward CFL metrics. We will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: Sec. 5.3 and Tab. 3: The checkpoint is selected based on 'the lowest CFL percent error on the validation split of InFlux++ Real.' This selection criterion directly targets the CFL metric family (fx/fy recall) that is then reported as improved, while EPE recall@50px drops from 25.1% to 18.7% and cx recall@1% drops from 76.5% to 67.2% (Tab. 3, Combined). The claim that finetuning 'consistently improves focal length estimation' is over-scoped: the improvement is consistent only for the metric the checkpoint was selected to optimize, and the simultaneous degradation across all other intrinsics metrics is consistent with a trade-off induced by the selection criterion. Without multiple random seeds or variance estimates, it is unclear whether the +6% fx recall@10% gain is robust or could arise from checkpoint selection variance in a single run. The authors should either (a) report results from

    Authors: The referee is correct that selecting the checkpoint based on CFL percent error introduces a bias toward the CFL metric family, and that the simultaneous degradation in EPE and principal point metrics is consistent with this selection bias. We acknowledge this limitation and will address it in two ways. First, we will run the finetuning experiment with multiple random seeds (at least 3) and report mean and standard deviation for all metrics in Tab. 3, so that the reader can assess whether the CFL gains and EPE/principal-point degradations are robust across runs. Second, we will add an explicit discussion in Sec. 5.3 noting that the checkpoint selection criterion was chosen because CFL (i.e., focal length) is the primary parameter of interest for dynamic intrinsics, but that this choice means the reported CFL improvements cannot be interpreted as evidence of general improvement across all intrinsics metrics. We will also report the final-epoch checkpoint results alongside the CFL-selected checkpoint to give the reader an additional reference point that is not conditioned on the CFL metric. We agree that without variance estimates, the single-run +6% fx recall@10% gain cannot be claimed as robust, and we will soften the language accordingly. revision: yes

  2. Referee: Sec. 5.3, Tab. 3: The abstract states that finetuning 'consistently improves focal length estimation across both InFlux++ Real and InFlux.' However, Tab. 3 shows that on InFlux, EPE recall@50px drops from 34.1% to 27.9% and EPE recall@10px drops from 2.95% to 2.09%. While fx/fy recall does improve, the simultaneous and substantial degradation in EPE—a metric that directly measures the geometric quality of predicted intrinsics—undermines the framing of the result as a straightforward improvement. The paper should more carefully qualify the claim in the abstract and Sec. 5.3 to acknowledge that the improvement is specific to CFL metrics and comes at the cost of EPE and principal point accuracy, rather than presenting it as a general improvement in intrinsics estimation.

    Authors: We agree. The abstract and Sec. 5.3 currently frame the finetuning result as a general improvement in focal length estimation, but the data show that this improvement is specific to CFL percent error metrics and comes at the cost of degraded EPE and principal point accuracy. We will revise the abstract to state that finetuning improves CFL percent error metrics while noting the trade-off in EPE and principal point accuracy. Specifically, we will change the abstract sentence from 'Finetuning existing intrinsics prediction methods on InFlux++ Synth consistently improves focal length estimation across both InFlux++ Real and InFlux' to something like: 'Finetuning on InFlux++ Synth improves focal length percent error metrics on both InFlux++ Real and InFlux, though at the cost of degraded principal point and EPE accuracy, suggesting that synthetic supervision is a promising but incomplete direction for RGB-based intrinsics prediction.' We will similarly revise the corresponding claims in Sec. 5.3 and the contributions list in Sec. 1 to accurately reflect the scope of the improvement. We note that Sec. 5.3 already acknowledges the degradation in the body text ('it worsens recall for cx, cy, and EPE'), but the abstract and contributions list do not, and we will make them consistent. revision: yes

Circularity Check

0 steps flagged

No significant circularity found; prior works used as infrastructure, not as circular premises

full rationale

The paper's central claim—that finetuning AnyCalib on InFlux++ Synth improves focal length estimation—is supported by an external evaluation: AnyCalib (an external method) is finetuned on synthetic data whose ground truth comes from the rendering engine's known parameters, then evaluated on InFlux++ Real and InFlux benchmarks. The synthetic ground truth is generated from Blender's known camera parameters (thin lens equation, Eq. in Sec. 4.1.3), not fitted to the evaluation benchmarks. The three prior works from the same group (InFlux [12], Infinigen [19], Infinigen Indoors [20]) are used as infrastructure: InFlux provides the calibration methodology and benchmark, Infinigen provides the procedural scene generator and renderer. None of these are invoked as uniqueness theorems or ansätze that would force the conclusion. The checkpoint selection on CFL validation error (Sec. 5.3) is a methodological concern about selection bias and generalizability of the 'consistent improvement' claim, but it is not circularity: the checkpoint is selected on CFL error and evaluated on fx/fy recall (related but not identical metrics), and the evaluation is on held-out test data. The thin lens model (1/LFL = 1/CFL + 1/LTO) is a standard physics equation, not a self-citation. The self-citations are normal infrastructure reuse, not load-bearing circular premises. The derivation chain is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

No new physical entities, particles, or forces are introduced. The free parameters are dataset generation choices (bounds, thresholds, sampling distributions) rather than fitted constants. The axioms are standard domain assumptions from computer vision and optics. The thin lens equation is a standard physics model, not invented for this paper.

free parameters (4)
  • LFL bounds [8mm, 100mm] = [8, 100] mm
    Hand-set range for bounded random walk of lens focal length (Sec. 4.1.4). Not fitted to evaluation data.
  • Distortion sampling parameters (mu_CFL coefficients, sigma values) = mu_CFL=0.0021*CFL+16.84, sigma=50, etc.
    Hand-tuned coefficients for realistic distortion distribution (Appendix F). Authors state they are hand-tuned.
  • Dominant surface rejection thresholds = q_outdoor>=0.35 (indoor), >=0.20 (nature)
    Hand-set thresholds for camera view validation (Appendix B).
  • Finetuning hyperparameters = lr=7.5e-6, batch=48, 15 epochs
    Standard hyperparameters for finetuning, not fitted to test set. Validation split used for checkpoint selection.
axioms (4)
  • domain assumption Blender's Cycles thin lens model accurately simulates real camera defocus blur and ray bending
    Sec. 4.1.3 and Appendix C. The entire synthetic dataset's realism depends on this assumption.
  • domain assumption Per-frame LFL and FD lens metadata uniquely determine the optical state of the lens
    Sec. 4.2, following InFlux [12]. Ground truth intrinsics for real data depend on this.
  • domain assumption LUT interpolation between calibrated (LFL, FD) grid points yields accurate intrinsics for intermediate values
    Sec. 5.1. The LUT-reliable EPE metric partially addresses this, but the assumption remains load-bearing for all real benchmark ground truth.
  • standard math Brown-Conrady radial distortion model adequately models lens distortion for the cameras used
    Sec. 4.1.5. Standard model but may not capture all real lens distortions.

pith-pipeline@v1.1.0-glm · 31282 in / 2215 out tokens · 287335 ms · 2026-07-07T12:37:51.353383+00:00 · methodology

0 comments
read the original abstract

Camera intrinsics are vital for recovering 3D structure from 2D video. However, most 3D algorithms assume fixed intrinsics throughout a video, an assumption that often fails for real-world in-the-wild videos. Consequently, estimating per-frame intrinsics from RGB images is critical for making 3D methods robust to videos with dynamic intrinsics. InFlux previously advanced this research direction by establishing the first real-world benchmark with per-frame ground truth intrinsics for dynamic intrinsics videos. Nevertheless, existing methods remain inaccurate due to two obstacles: (i) training data is scarce and lacks intrinsics diversity; and (ii) benchmarks, including InFlux, have limited scene and camera motion diversity, making it difficult to properly evaluate methods. To address both gaps, we present InFlux++, consisting of two components. InFlux++ Synth is a large-scale procedurally generated synthetic video dataset with 441K+ annotated frames from 1841 high-resolution videos, providing accurate per-frame ground truth intrinsics for training dynamic intrinsics prediction models; a subset also includes per-frame pose, depth, and normals. The videos feature rich intrinsics diversity through changes in camera zoom and focus, as well as dynamic objects and realistic rendering effects such as lens distortion and defocus blur. InFlux++ Real is a large-scale real-world benchmark that extends InFlux with 514K+ newly captured frames across 334 high-resolution videos, spanning a wider range of scenes and camera motions. Finetuning existing intrinsics prediction methods on InFlux++ Synth consistently improves focal length estimation across both InFlux++ Real and InFlux, suggesting that synthetic supervision is promising for RGB-based intrinsics prediction. For the dataset, benchmark, code, videos, submission instructions, and live leaderboard, please visit https://influx.cs.princeton.edu/ .

Figures

Figures reproduced from arXiv: 2607.05389 by Benjamin Zhou, Caleb Kha-Uong, Chinmaya Saran, David W. Liu, Erich Liang, Jia Deng, Junhan Ouyang, Sreemanti Dey.

Figure 1
Figure 1. Figure 1: A gallery of InFlux++, our data suite for dynamic intrinsics prediction. The top three rows show videos from InFlux++ Synth, a synthetic dataset with 441K+ frames across 1841 videos, featuring diverse intrinsics from changing zoom and focus, varying lens distortion, dynamic objects, and realistic optical effects such as defocus blur. The bottom three rows show videos from InFlux++ Real, a real-world dynami… view at source ↗
Figure 2
Figure 2. Figure 2: Gallery of InFlux++ Synth’s scene diversity. The dataset includes a wide range of procedurally generated indoor and nature environments with diverse lighting, ma￾terials, and scene composition. Indoor scenes feature varied room structures and non￾Lambertian surfaces, while nature scenes span settings like forests, deserts, mountains, and underwater environments with rich terrain, vegetation, water bodies, … view at source ↗
Figure 3
Figure 3. Figure 3: Gallery of InFlux++ Real’s scene diversity. Our benchmark contains a wide range of environments, activities, and viewpoints. vary the intrinsics across frames or videos. Among the few that do, [28] varies focal length but does not model lens distortion or defocus blur, limiting its usefulness for training models for dynamic intrinsics video prediction. The second obstacle to improving dynamic intrinsics pr… view at source ↗
Figure 4
Figure 4. Figure 4: Histograms of per-frame intrinsics in InFlux++ Synth. Values of k1 and k2 are clipped at the 0.1% and 99.9% percentiles for visualization. Both are heavily centered around 0. LTO ranges roughly from 0.04 m to 12 m for indoor scenes and from 0.7 m upward for nature scenes. These value ranges arise naturally from scene geometry. maximum visible scene depths at each keyframe via ray tracing, denoted as dnear … view at source ↗
Figure 5
Figure 5. Figure 5: LUT-reliable EPE recall plots for baseline methods, evaluated on the test splits of InFlux++ Real and InFlux for every integer threshold between 1 px and 300 px. AnyCalib performs the best, but still has low EPE recall for low pixel thresholds [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages

  1. [1]

    Atienza, R., Zelinsky, A.: A practical zoom camera calibration technique: An ap- plication of active vision for human-robot interaction (10 2002)

    work page 2002

  2. [2]

    Photogrammetric Engineering 32(3), 444–462 (1966)

    Brown, D.C.: Decentering Distortion of Lenses. Photogrammetric Engineering 32(3), 444–462 (1966)

    work page 1966

  3. [3]

    Burri, M., Nikolic, J., Gohl, P., Schneider, T., Rehder, J., Omari, S., Achtelik, M.W., Siegwart, R.: The euroc micro aerial vehicle datasets. Int. J. Rob. Res. 35(10), 1157–1163 (Sep 2016).https://doi.org/10.1177/0278364915620033, https://doi.org/10.1177/0278364915620033

    work page doi:10.1177/0278364915620033 2016

  4. [4]

    In: Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) (June 2020)

    Deng, J., Guo, J., Ververas, E., Kotsia, I., Zafeiriou, S.: Retinaface: Single-shot multi-level face localisation in the wild. In: Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) (June 2020)

    work page 2020

  5. [5]

    Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. Int. J. Rob. Res.32(11), 1231–1237 (Sep 2013).https://doi.org/10. 1177/0278364913491297,https://doi.org/10.1177/0278364913491297

    work page doi:10.1177/0278364913491297 2013

  6. [6]

    Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross B

    Hagemann, A., Knorr, M., Stiller, C.: Deep geometry-aware camera self-calibration from video. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3415–3425 (2023).https://doi.org/10.1109/ICCV51070.2023. 00318

    work page doi:10.1109/iccv51070.2023 2023

  7. [7]

    VideoMAE V2: Scaling Video Masked Autoencoders with Dual Masking

    Jin, L., Zhang, J., Hold-Geoffroy, Y., Wang, O., Blackburn-Matzen, K., Sticha, M., Fouhey, D.F.: Perspective Fields for Single Image Camera Calibration . In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17307–17316. IEEE Computer Society, Los Alamitos, CA, USA (Jun 2023).https://doi.org/10.1109/CVPR52729.2023.01660,https...

    work page doi:10.1109/cvpr52729.2023.01660 2023

  8. [8]

    ACM Transactions on Graphics42(4) (July 2023),https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics42(4) (July 2023),https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/

    work page 2023

  9. [9]

    In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision (ICCV) (2025)

    Krishnan, A., Liu, S., Sarlin, P.E., Gentilhomme, O., Caruso, D., Monge, M., New- combe, R., Engel, J., Pollefeys, M.: Benchmarking egocentric visual-inertial slam at city scale. In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision (ICCV) (2025)

    work page 2025

  10. [10]

    In: British Machine Vision Conference (BMVC) (2018)

    Li, W., Saeedi, S., McCormac, J., Clark, R., Tzoumanikas, D., Ye, Q., Huang, Y., Tang, R., Leutenegger, S.: Interiornet: Mega-scale multi-sensor photo-realistic indoor scenes dataset. In: British Machine Vision Conference (BMVC) (2018)

    work page 2018

  11. [11]

    In: Computer Vision and Pattern Recognition (CVPR) (2018)

    Li, Z., Snavely, N.: Megadepth: Learning single-view depth prediction from internet photos. In: Computer Vision and Pattern Recognition (CVPR) (2018)

    work page 2018

  12. [12]

    In: Belgrave, D., Zhang, C., Lin, H., Pascanu, R., Koniusz, P., Ghassemi, M., Chen, N

    Liang, E., Bhattacharjee, R., Dey, S., Moschopoulos, R., Wang, C., Liao, M., Tan, G., Wang, A., Kayan, K., Alexandropoulos, S., Deng, J.: Influx: A bench- mark for self-calibration of dynamic intrinsics of video cameras. In: Belgrave, D., Zhang, C., Lin, H., Pascanu, R., Koniusz, P., Ghassemi, M., Chen, N. (eds.) Advances in Neural Information Processing ...

    work page 2025

  13. [13]

    McCormac, J., Handa, A., Leutenegger, S., Davison, A.J.: Scenenet rgb-d: Can 5m synthetic images beat generic imagenet pre-training on indoor segmentation? In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (Oct 2017) InFlux++ 17

    work page 2017

  14. [14]

    In: Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I. p. 405–421. Springer-Verlag, Berlin, Heidelberg (2020).https://doi.org/10.1007/978-3-03...

    work page doi:10.1007/978-3-030-58452-8_24 2020

  15. [15]

    IEEE Transactions on Robotics33(5), 1255–1262 (2017).https://doi.org/10.1109/TRO.2017.2705103

    Mur-Artal, R., Tardós, J.D.: ORB-SLAM2: an open-source SLAM system for monocular, stereo and RGB-D cameras. IEEE Transactions on Robotics33(5), 1255–1262 (2017).https://doi.org/10.1109/TRO.2017.2705103

    work page doi:10.1109/tro.2017.2705103 2017

  16. [16]

    In: Proceedings of the IEEE International Conference on Robotics and Automation (ICRA)

    Olson, E.: AprilTag: A robust and flexible visual fiducial system. In: Proceedings of the IEEE International Conference on Robotics and Automation (ICRA). pp. 3400–3407. IEEE (May 2011)

    work page 2011

  17. [17]

    IEEE Transactions on Pattern Analysis and Machine Intelligence48(3), 2354–2367 (2026).https://doi.org/10.1109/TPAMI.2025.3628473

    Piccinelli, L., Sakaridis, C., Yang, Y.H., Segu, M., Li, S., Abbeloos, W., Van Gool, L.: Unidepthv2: Universal monocular metric depth estimation made simpler. IEEE Transactions on Pattern Analysis and Machine Intelligence48(3), 2354–2367 (2026).https://doi.org/10.1109/TPAMI.2025.3628473

    work page doi:10.1109/tpami.2025.3628473 2026

  18. [18]

    Raina, N., Somasundaram, G., Zheng, K., Miglani, S., Saarinen, S., Meissner, J., Schwesinger, M., Pesqueira, L., Prasad, I., Miller, E., Gupta, P., Yan, M., New- combe, R., Ren, C., Parkhi, O.M.: Egoblur: Responsible innovation in aria (2023)

    work page 2023

  19. [19]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Raistrick, A., Lipson, L., Ma, Z., Mei, L., Wang, M., Zuo, Y., Kayan, K., Wen, H., Han, B., Wang, Y., Newell, A., Law, H., Goyal, A., Yang, K., Deng, J.: Infinite pho- torealistic worlds using procedural generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12630–12641 (2023)

    work page 2023

  20. [20]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR)

    Raistrick, A., Mei, L., Kayan, K., Yan, D., Zuo, Y., Han, B., Wen, H., Parakh, M., Alexandropoulos, S., Lipson, L., Ma, Z., Deng, J.: Infinigen indoors: Photorealistic indoor scenes using procedural generation. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR). pp. 21783–21794 (June 2024)

    work page 2024

  21. [21]

    Mastering Photography pp

    Roberts, G.: Lenses. Mastering Photography pp. 41–63 (1995)

    work page 1995

  22. [22]

    In: ECCV (2022)

    Sarlin, P.E., Dusmanu, M., Schönberger, J.L., Speciale, P., Gruber, L., Larsson, V., Miksik, O., Pollefeys, M.: LaMAR: Benchmarking Localization and Mapping for Augmented Reality. In: ECCV (2022)

    work page 2022

  23. [23]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2016)

    Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2016)

    work page 2016

  24. [24]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (July 2017)

    Schops, T., Schonberger, J.L., Galliani, S., Sattler, T., Schindler, K., Pollefeys, M., Geiger, A.: A multi-view stereo benchmark with high-resolution images and multi-camera videos. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (July 2017)

    work page 2017

  25. [25]

    In: 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems

    Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of rgb-d slam systems. In: 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems. pp. 573–580 (2012).https://doi.org/10. 1109/IROS.2012.6385773

    work page arXiv 2012

  26. [26]

    In: Proceedings of the 35th International Conference on Neural Informa- tion Processing Systems

    Teed, Z., Deng, J.: Droid-slam: deep visual slam for monocular, stereo, and rgb-d cameras. In: Proceedings of the 35th International Conference on Neural Informa- tion Processing Systems. NIPS ’21, Curran Associates Inc., Red Hook, NY, USA (2021)

    work page 2021

  27. [27]

    Advances in Neural Information Processing Systems36, 39033–39051 (2023), https : / / proceedings

    Teed, Z., Lipson, L., Deng, J.: Deep patch visual odometry. Advances in Neural Information Processing Systems36, 39033–39051 (2023), https : / / proceedings . neurips . cc / paper _ files / paper / 2023 / file / 7ac484b0f1a1719ad5be9aa8c8455fbb-Paper-Conference.pdf 18 E. Liang et al

    work page 2023

  28. [28]

    In: The Thirty- ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2025)

    Tesch, J., Becherini, G., Achar, P., Yiannakidis, A., Kocabas, M., Patel, P., Black, M.J.: BEDLAM2.0: Synthetic humans and cameras in motion. In: The Thirty- ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2025)

    work page 2025

  29. [29]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Tirado-Garín, J., Civera, J.: Anycalib: On-manifold learning for model-agnostic single-view camera calibration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8044–8055 (October 2025)

    work page 2025

  30. [30]

    In: Computer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29–October 4, 2024, Pro- ceedings, Part XL

    Veicht, A., Sarlin, P.E., Lindenberger, P., Pollefeys, M.: Geocalib: Learning single- image calibration with geometric optimization. In: Computer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29–October 4, 2024, Pro- ceedings, Part XL. p. 1–20. Springer-Verlag, Berlin, Heidelberg (2024).https: //doi.org/10.1007/978-3-031-73661-2_1,...

    work page doi:10.1007/978-3-031-73661-2_1 2024

  31. [31]

    In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

    Wang, W., Zhu, D., Wang, X., Hu, Y., Qiu, Y., Wang, C., Hu, Y., Kapoor, A., Scherer, S.: Tartanair: A dataset to push the limits of visual slam. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). p. 4909–4916. IEEE Press (2020).https://doi.org/10.1109/IROS45743.2020. 9341801,https://doi.org/10.1109/IROS45743.2020.9341801

    work page doi:10.1109/iros45743.2020 2020

  32. [32]

    #$%&'(𝐴⃗,𝐵=𝑑!

    Zhu, S., Kumar, A., Hu, M., Liu, X.: Tame a wild camera: in-the-wild monocular camera calibration. In: Proceedings of the 37th International Conference on Neural Information Processing Systems. NIPS ’23, Curran Associates Inc., Red Hook, NY, USA (2023) InFlux++ 19 Appendix A Additional Details on InFlux++ Synth: Camera Pose, Depth, Surface Normals, and Ot...

    work page 2023