pith. sign in

arxiv: 2606.03989 · v1 · pith:XMCOWVILnew · submitted 2026-06-02 · 💻 cs.CV

PixVOD: Pixel-Distributed Direct Visual Odometry and Depth Estimation

Shinjeong Kim , Ignacio Alzugaray , Callum Rhodes , Paul H. J. Kelly , Andrew J. Davison This is my paper

Pith reviewed 2026-06-28 10:13 UTC · model grok-4.3

classification 💻 cs.CV
keywords visual odometrydepth estimationGaussian Belief Propagationfocal-plane sensor-processorspixel-distributed computationdirect visual odometrysurface normal priorkeyframe anchoring
0
0 comments X

The pith

Pixels can estimate camera motion and scene depth by exchanging Gaussian beliefs locally on the sensor.

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

The paper proposes distributing visual odometry and depth estimation across individual pixels on focal-plane sensor-processors. Each pixel contributes photometric observations and receives messages via Gaussian Belief Propagation to reach consensus on camera motion while using a surface normal prior to infer depth. A keyframe-like anchoring mechanism regulates the baseline between frames to keep the optimization geometrically stable. The goal is to let pixels synthesize higher-level signals on-sensor rather than transmitting raw pixel arrays, reducing data movement for downstream vision tasks.

Core claim

Gaussian Belief Propagation enables a fully parallelizable form of visual odometry and depth estimation across pixels, where sensor-processors exchange information to achieve consensus about camera motion and infer depth from per-pixel photometric observations and a surface normal prior, with keyframe-like anchoring regulating the effective baseline to maintain geometric stability.

What carries the argument

Gaussian Belief Propagation (GBP) messages exchanged between pixels for consensus on motion and depth, together with keyframe-like anchoring to control baseline.

Load-bearing premise

Gaussian Belief Propagation messages can be exchanged and converged efficiently inside focal-plane sensor-processors, and the surface normal prior plus photometric observations suffice to produce usable depth and motion estimates without external calibration or additional sensors.

What would settle it

A focal-plane hardware implementation where GBP messages fail to converge within frame timing or yield depth and motion estimates that diverge from ground truth on standard datasets.

Figures

Figures reproduced from arXiv: 2606.03989 by Andrew J. Davison, Callum Rhodes, Ignacio Alzugaray, Paul H. J. Kelly, Shinjeong Kim.

Figure 1
Figure 1. Figure 1: Most vision algorithms assume the use of an off-sensor central processor with access to all pixel information. In contrast, we present an algorithm to track camera pose and estimate depth on the assumption that each pixel has its own local processing capability, with only local memory and the ability to exchange messages with other pixels. Our goal is to motivate and guide the design of near-future chips a… view at source ↗
Figure 2
Figure 2. Figure 2: Topology of the proposed factor graph. The factor graph on the right side of the Image layer, related to camera pose estimation, follows the sharded topology of PixRO [3], with the domain extended to SE(3). The factor graph on the left side of the Image layer reconstructs depth through the normal integration and photometric factors. For the normal integration factors, note that all factors except the one l… view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of an estimated local trajectory and depth with (top) relatively small motion and being close to surface, (bottom) relatively large motion and being far from surface. The view frustum of the camera is visualized only on the final pose. For the top image, the trajectory from a reference keyframe to the next one is shown, and for the bottom trajectory, 5 keyframes are in between, while omitted … view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of estimated pose error across various frame rates, based on the deviations of the target camera pose from origin [I; 0]. As the pose change between consecutive frames increases, i.e., as the frame rate decreases, the convergence basin of the proposed method (with respect to the deviation of the target camera pose from the origin) becomes smaller, and the corresponding error increases. However, … view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of the proposed method against the baseline method, which is solved in a centralized manner. We formulated the centralized optimization problem as an Iterative Reweighted Least Squares (IRLS), and solved it with the Gauss-Newton method. 4.4 Comparison of Pixel-Parallel and Centralized Visual Odometry Estimation We compare the optimization performance of the proposed GBP-based method against the … view at source ↗
Figure 6
Figure 6. Figure 6: (a) Comparison of reconstructed depth of BINI [11] after convergence against our reconstructed depth with various translational baselines of camera. (b) Evaluation on camera poses estimated with various structural priors. Note that most of the exper￾iments reported in this paper are paired with the Ground Truth Normals. for the BINI results. Consistent with Sec. 4.3, there exists a range of camera translat… view at source ↗
Figure 7
Figure 7. Figure 7: (Left) [PITH_FULL_IMAGE:figures/full_fig_p022_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Error of estimated poses across various strengths of the camera motion identity factor. were possible, it would introduce a severe inter-pixel communication bottleneck. This constitutes an additional practical challenge that is not reflected in the FLOP counts reported in our table. D.5 Scalability The proposed pixel-distributed optimization framework has the advantage that the computational workload assig… view at source ↗
Figure 9
Figure 9. Figure 9: Error of estimated poses across various strategies of assigning camera pose identity and prior strengths on different layers of sharded factor graph topology. 500 1500 2500 3500 4500 5500 6500 7500 8500 9500 Iteration 0.0 0.5 1.0 1.5 2.0 Translation Relative Error 500 1500 2500 3500 4500 5500 6500 7500 8500 9500 Iteration 0.0 0.5 1.0 1.5 2.0 Rotation Relative Error Prior Strength: Stronger ( =0.2) PixVOD (… view at source ↗
Figure 10
Figure 10. Figure 10: Error of estimated poses across various strengths of the prior factor. optimization to converge to inaccurate camera pose estimates. Conversely, when the camera motion identity factor is too strong, the photometric factors—which are the primary source of informative constraints in our system—are effectively downweighted, and the optimization again converges to inaccurate camera poses. E.2 Inter-layer Diff… view at source ↗
Figure 11
Figure 11. Figure 11: Error of estimated poses across various numbers of iterations to optimize per frame. 1000 3000 5000 7000 9000 11000 13000 15000 17000 19000 Iteration 0.0 0.5 1.0 1.5 2.0 2.5 Translation Relative Error 1000 3000 5000 7000 9000 11000 13000 15000 17000 19000 Iteration 0.0 0.5 1.0 1.5 2.0 2.5 Rotation Relative Error Resolution: (32, 32) (64, 64) (128, 128) [PITH_FULL_IMAGE:figures/full_fig_p027_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Error of estimated poses across various resolutions of input images. E.3 Prior Strength As shown in [PITH_FULL_IMAGE:figures/full_fig_p027_12.png] view at source ↗
read the original abstract

Images composed of 2D pixel arrays are the standard input to computer vision algorithms, yet many underlying computations can be distributed across pixels. Transmitting raw, redundant, and noisy pixel data off the sensor remains inefficient, motivating a shift toward focal-plane sensor-processors that perform a significant part of the computation directly within each pixel. We envision pixels synthesizing higher-level signals locally, reducing downstream load, and providing richer inputs for higher-level vision tasks. We propose a fully parallelizable form of visual odometry and depth estimation across pixels, where sensor-processors exchange information through Gaussian Belief Propagation (GBP) to achieve consensus about camera motion and infer depth from per-pixel photometric observations and a surface normal prior. To maintain geometric stability during optimization, we introduce a keyframe-like anchoring mechanism that regulates the effective baseline between frames, enabling consistent motion and depth updates. Our method is evaluated on realistic datasets, demonstrating the feasibility of GBP-based pixel-level distributed odometry and depth estimation with keyframe anchoring on-sensor. Project Page: https://www.shinjeongkim.com/pixvod/

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes PixVOD, a fully parallelizable pixel-distributed direct visual odometry and depth estimation algorithm in which focal-plane sensor-processors exchange Gaussian Belief Propagation (GBP) messages to reach consensus on camera motion while inferring per-pixel depth from photometric observations and a surface normal prior. A keyframe-like anchoring mechanism is introduced to regulate effective baseline and maintain geometric stability. The method is evaluated on realistic datasets and presented as demonstrating feasibility of GBP-based on-sensor odometry and depth estimation.

Significance. If the distributed GBP formulation and anchoring mechanism can be shown to converge under realistic focal-plane constraints, the work would offer a concrete route toward reducing raw pixel data transmission by performing VO and depth inference locally. The simulation results on standard datasets provide an initial existence proof for the algorithmic construction, but the hardware-specific claims remain untested.

major comments (2)
  1. [Evaluation section] Evaluation section (and abstract): the central claim that GBP message exchange enables on-sensor consensus about global camera motion is supported only by conventional software simulation on standard datasets; no experiments or analysis quantify communication topology, message volume per iteration, iteration count to convergence, bandwidth, or power limits of actual focal-plane sensor-processor arrays.
  2. [Method description] Method description (surface normal prior): the surface normal prior is taken as given per pixel, yet neither its acquisition cost nor its accuracy sensitivity is quantified with respect to the distributed GBP factors; this directly affects whether photometric observations plus the prior suffice without external calibration.
minor comments (1)
  1. [Abstract] The abstract and introduction would benefit from an explicit statement of the assumed sensor-processor architecture (e.g., interconnect topology and per-pixel compute budget) to make the hardware feasibility claim concrete.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. The comments highlight important aspects of scope and assumptions in our simulation-based work. We respond point-by-point below and indicate planned revisions.

read point-by-point responses

  1. Referee: [Evaluation section] Evaluation section (and abstract): the central claim that GBP message exchange enables on-sensor consensus about global camera motion is supported only by conventional software simulation on standard datasets; no experiments or analysis quantify communication topology, message volume per iteration, iteration count to convergence, bandwidth, or power limits of actual focal-plane sensor-processor arrays.

    Authors: We agree the evaluation uses conventional software simulation on standard datasets and does not report hardware-specific metrics such as power or bandwidth on physical focal-plane arrays. The manuscript presents an algorithmic construction and existence proof for distributed GBP-based VO and depth estimation. We will revise the abstract and evaluation section to explicitly qualify the simulation setting. We will also add analysis of the simulated GBP process, including iteration counts to convergence, per-iteration message volume, and communication topology statistics derived from the pixel grid. Direct hardware measurements remain outside the paper's scope. revision: partial

  2. Referee: [Method description] Method description (surface normal prior): the surface normal prior is taken as given per pixel, yet neither its acquisition cost nor its accuracy sensitivity is quantified with respect to the distributed GBP factors; this directly affects whether photometric observations plus the prior suffice without external calibration.

    Authors: The surface normal prior is supplied as a per-pixel factor in the GBP model. We will add a sensitivity study in the revised method and evaluation sections that perturbs the prior normals and reports resulting changes in depth and motion accuracy. We will also include a brief discussion of acquisition approaches (e.g., on-sensor estimation or auxiliary modalities) and their relation to the photometric factors. revision: yes

Circularity Check

0 steps flagged

No circularity: new algorithmic construction with independent evaluation

full rationale

The paper introduces a distributed GBP-based visual odometry and depth estimation method across pixels, with a novel keyframe-like anchoring mechanism for stability. The abstract and description present this as an original algorithmic proposal evaluated on realistic datasets, without any equations or steps that reduce predictions to fitted inputs by construction, self-definitional loops, or load-bearing self-citations that collapse the central claim. The surface normal prior and GBP consensus are treated as modeling choices with external validation via simulation results, keeping the derivation self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The proposal rests on the standard properties of Gaussian Belief Propagation and the assumption that a surface normal prior plus photometry is adequate; no free parameters or invented entities are mentioned in the abstract.

axioms (2)
  • domain assumption Gaussian Belief Propagation messages can be passed and converged on focal-plane hardware to reach consensus on camera motion and depth
    Central mechanism invoked in the abstract
  • domain assumption Surface normal prior combined with per-pixel photometric observations is sufficient to infer depth
    Stated as the basis for depth inference

pith-pipeline@v0.9.1-grok · 5733 in / 1132 out tokens · 20187 ms · 2026-06-28T10:13:17.559669+00:00 · methodology

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

72 extracted references · 9 canonical work pages · 1 internal anchor

  1. [1]

    Communications of the ACM54(10), 105–112 (2011)

    Agarwal, S., Furukawa, Y., Snavely, N., Simon, I., Curless, B., Seitz, S.M., Szeliski, R.: Building rome in a day. Communications of the ACM54(10), 105–112 (2011)

    2011

  2. [2]

    In: Fourth Swedish Workshop on Computer Systems Architecture (1992)

    Ahlander, A., Svensson, B.: Floating point calculations in bit-serial simd comput- ers. In: Fourth Swedish Workshop on Computer Systems Architecture (1992)

    1992

  3. [3]

    arXiv preprint arXiv:2406.09726 (2024)

    Alzugaray, I., Murai, R., Davison, A.: Pixro: Pixel-distributed rotational odometry with gaussian belief propagation. arXiv preprint arXiv:2406.09726 (2024)

    work page arXiv 2024

  4. [4]

    In: 2020 57th ACM/IEEE Design Automation Conference (DAC)

    Asgari, B., Hadidi, R., Ghaleshahi, N.S., Kim, H.: Pisces: power-aware implemen- tation of slam by customizing efficient sparse algebra. In: 2020 57th ACM/IEEE Design Automation Conference (DAC). pp. 1–6. IEEE (2020)

    2020

  5. [5]

    In: IEEE Conf

    Bae, G., Davison, A.J.: Rethinking inductive biases for surface normal estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 9535–9545 (2024)

    2024

  6. [6]

    Bishop, C.M., Nasrabadi, N.M.: Pattern recognition and machine learning, vol. 4. Springer (2006)

    2006

  7. [7]

    In: Yue, Y., Garg, A., Peng, N., Sha, F., Yu, R

    Bochkovskiy, A., Delaunoy, A., Germain, H., Santos, M., Zhou, Y., Richter, S., Koltun, V.: Depth pro: Sharp monocular metric depth in less than a second. In: Yue, Y., Garg, A., Peng, N., Sha, F., Yu, R. (eds.) Int. Conf. Learn. Represent. vol. 2025, pp. 75602–75637 (2025),https://proceedings.iclr.cc/paper_files/ paper/2025/file/bc8b2058fd96978a4146f18298c...

    2025

  8. [8]

    In: 2016 26th Interna- tional Conference on Field Programmable Logic and Applications (FPL)

    Boikos, K., Bouganis, C.S.: Semi-dense slam on an fpga soc. In: 2016 26th Interna- tional Conference on Field Programmable Logic and Applications (FPL). pp. 1–4. IEEE (2016)

    2016

  9. [9]

    In: IEEE Conf

    Bose, L., Chen, J., Dudek, P.: Descriptor-in-pixel: Point-feature tracking for pixel processor arrays. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5392–5400 (2025)

    2025

  10. [10]

    In: IEEE Conf

    Bose, L., Dudek, P., Carey, S.J., Chen, J.: Live demonstration: Scamp-7. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3995–3996 (2023)

    2023

  11. [11]

    Cao, X., Santo, H., Shi, B., Okura, F., Matsushita, Y.: Bilateral normal integration. In: Eur. Conf. Comput. Vis. pp. 552–567. Springer (2022)

    2022

  12. [12]

    In: IEEE Conf

    Cao, X., Shi, B., Okura, F., Matsushita, Y.: Normal integration via inverse plane fitting with minimum point-to-plane distance. In: IEEE Conf. Comput. Vis. Pat- tern Recog. pp. 2382–2391 (2021)

    2021

  13. [13]

    In: 2013 symposium on VLSI circuits

    Carey, S.J., Lopich, A., Barr, D.R., Wang, B., Dudek, P.: A 100,000 fps vision sen- sor with embedded 535gops/w 256×256 simd processor array. In: 2013 symposium on VLSI circuits. pp. C182–C183. IEEE (2013)

    2013

  14. [14]

    Chatterjee, A., Govindu, V.M.: Efficient and robust large-scale rotation averaging. In: Int. Conf. Comput. Vis. pp. 521–528 (2013)

    2013

  15. [15]

    In: IEEE Conf

    Dai,A.,Chang,A.X.,Savva,M.,Halber,M.,Funkhouser,T.,Nießner,M.:Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5828–5839 (2017)

    2017

  16. [16]

    arXiv preprint arXiv:1910.14139 (2019) 16 S

    Davison, A.J., Ortiz, J.: Futuremapping 2: Gaussian belief propagation for spatial ai. arXiv preprint arXiv:1910.14139 (2019) 16 S. Kim et al

    work page arXiv 1910

  17. [17]

    Foundations and Trends®in Robotics6(1-2), 1–139 (2017)

    Dellaert, F., Kaess, M., et al.: Factor graphs for robot perception. Foundations and Trends®in Robotics6(1-2), 1–139 (2017)

    2017

  18. [18]

    In: IEEE Conf

    DeTone, D., Malisiewicz, T., Rabinovich, A.: Superpoint: Self-supervised inter- est point detection and description. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. pp. 224–236 (2018)

    2018

  19. [19]

    IEEE Trans

    Engel, J., Koltun, V., Cremers, D.: Direct sparse odometry. IEEE Trans. Pattern Anal. Mach. Intell.40(3), 611–625 (2017)

    2017

  20. [20]

    In: 2017 International Conference on Field Programmable Technology (ICFPT)

    Fang,W.,Zhang,Y.,Yu,B.,Liu,S.:Fpga-basedorbfeatureextractionforreal-time visual slam. In: 2017 International Conference on Field Programmable Technology (ICFPT). pp. 275–278. IEEE (2017)

    2017

  21. [21]

    IEEE Transactions on Im- age Processing17(10), 1737–1754 (2008).https://doi.org/10.1109/TIP.2008

    Foi, A., Trimeche, M., Katkovnik, V., Egiazarian, K.: Practical poissonian-gaussian noise modeling and fitting for single-image raw-data. IEEE Transactions on Im- age Processing17(10), 1737–1754 (2008).https://doi.org/10.1109/TIP.2008. 2001399

    work page doi:10.1109/tip.2008 2008

  22. [22]

    In: IEEE Conf

    Guan, T., Wang, C., Liu, Y.H.: Neural markov random field for stereo matching. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5459–5469 (2024)

    2024

  23. [23]

    Handa, A., Newcombe, R.A., Angeli, A., Davison, A.J.: Real-time camera tracking: When is high frame-rate best? In: Eur. Conf. Comput. Vis. pp. 222–235. Springer (2012)

    2012

  24. [24]

    Heep, M., Zell, E.: An adaptive screen-space meshing approach for normal integra- tion. In: Eur. Conf. Comput. Vis. pp. 445–461. Springer (2024)

    2024

  25. [25]

    International Journal of Computer Vision129(2), 517–547 (2021)

    Jin, Y., Mishkin, D., Mishchuk, A., Matas, J., Fua, P., Yi, K.M., Trulls, E.: Image matching across wide baselines: From paper to practice. International Journal of Computer Vision129(2), 517–547 (2021)

    2021

  26. [26]

    Trinity College Dublin5, 4 (2006)

    Kavan, L., Collins, S., O’Sullivan, C., Zara, J.: Dual quaternions for rigid trans- formation blending. Trinity College Dublin5, 4 (2006)

    2006

  27. [27]

    In: IEEE Conf

    Kim, H., Jung, Y., Lee, S.: Discontinuity-preserving normal integration with aux- iliary edges. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11915–11923 (2024)

    2024

  28. [28]

    Krähenbühl, P., Koltun, V.: Efficient inference in fully connected crfs with gaussian edge potentials. Adv. Neural Inform. Process. Syst.24(2011)

    2011

  29. [29]

    IEEE Transactions on Circuits and Systems for Video Technology26(8), 1563–1577 (2015)

    Lentaris, G., Stamoulias, I., Soudris, D., Lourakis, M.: Hw/sw codesign and fpga acceleration of visual odometry algorithms for rover navigation on mars. IEEE Transactions on Circuits and Systems for Video Technology26(8), 1563–1577 (2015)

    2015

  30. [30]

    arXiv preprint arXiv:2505.04612 (2025)

    Li, J., Wang, H., Irshad, M.Z., Vasiljevic, I., Walter, M.R., Guizilini, V.C., Shakhnarovich, G.: Fastmap: Revisiting structure from motion through first-order optimization. arXiv preprint arXiv:2505.04612 (2025)

    work page arXiv 2025

  31. [31]

    Lindenberger, P., Sarlin, P.E., Pollefeys, M.: Lightglue: Local feature matching at light speed. In: Int. Conf. Comput. Vis. pp. 17627–17638 (2023)

    2023

  32. [32]

    In: IEEE Conf

    Liu, S., Yu, Y., Pautrat, R., Pollefeys, M., Larsson, V.: 3d line mapping revisited. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21445–21455 (2023)

    2023

  33. [33]

    Frontiers in Neuroscience16(2022)

    Liu, Y., Bose, L., Fan, R., Dudek, P., Mayol-Cuevas, W.: On-sensor binarized CNN inference with dynamic model swapping in pixel processor arrays. Frontiers in Neuroscience16(2022)

    2022

  34. [34]

    https://www.graphcore.ai/products/ipu

    Ltd, G.: IPU Processors. https://www.graphcore.ai/products/ipu

  35. [35]

    In: 2019 Design, Automation & Test in Europe Conference & Exhibition (DATE)

    Mandal, D.K., Jandhyala, S., Omer, O.J., Kalsi, G.S., George, B., Neela, G., Rethi- nagiri, S.K., Subramoney, S., Hacking, L., Radford, J., et al.: Visual inertial odom- etry at the edge: A hardware-software co-design approach for ultra-low latency and power. In: 2019 Design, Automation & Test in Europe Conference & Exhibition (DATE). pp. 960–963. IEEE (2...

    2019

  36. [36]

    In: IEEE Conf

    Mazur, K., Bae, G., Davison, A.J.: Superprimitive: Scene reconstruction at a prim- itive level. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4979–4989 (2024)

    2024

  37. [37]

    IEEE Robot

    Murai, R., Alzugaray, I., Kelly, P.H., Davison, A.J.: Distributed simultaneous lo- calisation and auto-calibration using gaussian belief propagation. IEEE Robot. and Auto. Letters9(3), 2136–2143 (2024)

    2024

  38. [38]

    IEEE Trans

    Murai, R., Ortiz, J., Saeedi, S., Kelly, P.H., Davison, A.J.: A robot web for dis- tributed many-device localization. IEEE Trans. on Robotics40, 121–138 (2023)

    2023

  39. [39]

    In: IEEE/RSJ Int

    Murai, R., Saeedi, S., Kelly, P.H.: Bit-vo: Visual odometry at 300 fps using bi- nary features from the focal plane. In: IEEE/RSJ Int. Conf. on Intell. Robots and Systems. pp. 8579–8586. IEEE (2020)

    2020

  40. [40]

    arXiv preprint arXiv:2311.14649 (2023)

    Nabarro, S., Van Der Wilk, M., Davison, A.J.: Learning in deep factor graphs with gaussian belief propagation. arXiv preprint arXiv:2311.14649 (2023)

    work page arXiv 2023

  41. [41]

    In: IEEE Conf

    Nagata, J., Sekikawa, Y.: Tangentially elongated gaussian belief propagation for event-based incremental optical flow estimation. In: IEEE Conf. Comput. Vis. Pat- tern Recog. pp. 21940–21949 (2023)

    2023

  42. [42]

    Newcombe, R.A., Lovegrove, S.J., Davison, A.J.: Dtam: Dense tracking and map- ping in real-time. In: Int. Conf. Comput. Vis. pp. 2320–2327. IEEE (2011)

    2011

  43. [43]

    In: IEEE Int

    Nikolic, J., Rehder, J., Burri, M., Gohl, P., Leutenegger, S., Furgale, P.T., Sieg- wart, R.: A synchronized visual-inertial sensor system with fpga pre-processing for accurate real-time slam. In: IEEE Int. Conf. Robot. Autom. pp. 431–437. IEEE (2014)

    2014

  44. [44]

    arXiv preprint arXiv:2107.02308 (2021)

    Ortiz, J., Evans, T., Davison, A.J.: A visual introduction to gaussian belief prop- agation. arXiv preprint arXiv:2107.02308 (2021)

    work page arXiv 2021

  45. [45]

    In: IEEE Conf

    Ortiz,J.,Pupilli,M.,Leutenegger,S.,Davison,A.J.:Bundleadjustmentonagraph processor. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2416–2425 (2020)

    2020

  46. [46]

    Pan, L., Baráth, D., Pollefeys, M., Schönberger, J.L.: Global structure-from-motion revisited. In: Eur. Conf. Comput. Vis. pp. 58–77. Springer (2024)

    2024

  47. [47]

    In: IEEE Conf

    Pataki, Z., Sarlin, P.E., Schönberger, J.L., Pollefeys, M.: Mp-sfm: Monocular sur- face priors for robust structure-from-motion. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21891–21901 (2025)

    2025

  48. [48]

    Elsevier (2014)

    Pearl, J.: Probabilistic reasoning in intelligent systems: networks of plausible in- ference. Elsevier (2014)

    2014

  49. [49]

    Neural Networks2(6), 475–494 (1989)

    Peterson, C., Hartman, E.: Explorations of the mean field theory learning algo- rithm. Neural Networks2(6), 475–494 (1989)

    1989

  50. [50]

    In: IEEE Conf

    Qi, X., Liao, R., Liu, Z., Urtasun, R., Jia, J.: Geonet: Geometric neural network for joint depth and surface normal estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 283–291 (2018)

    2018

  51. [51]

    IEEE Trans

    Qi, X., Liu, Z., Liao, R., Torr, P.H., Urtasun, R., Jia, J.: Geonet++: Iterative geometric neural network with edge-aware refinement for joint depth and surface normal estimation. IEEE Trans. Pattern Anal. Mach. Intell.44(2), 969–984 (2020)

    2020

  52. [52]

    In: 2019 IEEE 27th Annual International Symposium on Field-Programmable Custom Computing Machines (FCCM)

    Qin, S., Liu, Q., Yu, B., Liu, S.:π-ba: Bundle adjustment acceleration on embedded fpgas with co-observation optimization. In: 2019 IEEE 27th Annual International Symposium on Field-Programmable Custom Computing Machines (FCCM). pp. 100–108. IEEE (2019)

    2019

  53. [53]

    Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: Superglue: Learning featurematchingwithgraphneuralnetworks.In:IEEEConf.Comput.Vis.Pattern Recog. pp. 4938–4947 (2020)

    2020

  54. [54]

    In: IEEE Conf

    Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4104–4113 (2016) 18 S. Kim et al

    2016

  55. [55]

    IEEE Robot

    Scona, R., Matsuki, H., Davison, A.: From scene flow to visual odometry through local and global regularisation in markov random fields. IEEE Robot. and Auto. Letters7(2), 4299–4306 (2022)

    2022

  56. [56]

    In: Proceedings of the 12th annual conference on Computer graphics and interactive techniques

    Shoemake, K.: Animating rotation with quaternion curves. In: Proceedings of the 12th annual conference on Computer graphics and interactive techniques. pp. 245– 254 (1985)

    1985

  57. [57]

    In: IEEE Conf

    So, H.M., Bose, L., Dudek, P., Wetzstein, G.: Pixelrnn: in-pixel recurrent neural networks for end-to-end-optimized perception with neural sensors. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 25233–25244 (2024)

    2024

  58. [58]

    A micro lie theory for state estimation in robotics,

    Sola, J., Deray, J., Atchuthan, D.: A micro lie theory for state estimation in robotics. arXiv preprint arXiv:1812.01537 (2018)

    work page arXiv 2018

  59. [59]

    The Replica Dataset: A Digital Replica of Indoor Spaces

    Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)

    work page internal anchor Pith review Pith/arXiv arXiv 1906

  60. [60]

    Sucar, E., Liu, S., Ortiz, J., Davison, A.J.: imap: Implicit mapping and positioning in real-time. In: Int. Conf. Comput. Vis. pp. 6229–6238 (2021)

    2021

  61. [61]

    Teed, Z., Deng, J.: Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Adv. Neural Inform. Process. Syst.34, 16558–16569 (2021)

    2021

  62. [62]

    In: 2022 International Conference on Field-Programmable Technology (ICFPT)

    Vemulapati, V., Chen, D.: Fslam: an efficient and accurate slam accelerator on soc fpgas. In: 2022 International Conference on Field-Programmable Technology (ICFPT). pp. 1–9. IEEE (2022)

    2022

  63. [63]

    In: IEEE Conf

    Wang, R., Xu, S., Dai, C., Xiang, J., Deng, Y., Tong, X., Yang, J.: Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5261–5271 (2025)

    2025

  64. [64]

    Yang, L., Kang, B., Huang, Z., Zhao, Z., Xu, X., Feng, J., Zhao, H.: Depth anything v2. Adv. Neural Inform. Process. Syst.37, 21875–21911 (2024)

    2024

  65. [65]

    In: IEEE/RSJ Int

    Yi, B., Lee, M.A., Kloss, A., Martín-Martín, R., Bohg, J.: Differentiable factor graph optimization for learning smoothers. In: IEEE/RSJ Int. Conf. on Intell. Robots and Systems. pp. 1339–1345. IEEE (2021)

    2021

  66. [66]

    In: IEEE Conf

    Yu, Y., Liu, S., Pautrat, R., Pollefeys, M., Larsson, V.: Relative pose estimation through affine corrections of monocular depth priors. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 16706–16716 (2025)

    2025

  67. [67]

    In: IEEE Conf

    Yuan, W., Gu, X., Dai, Z., Zhu, S., Tan, P.: Neural window fully-connected crfs for monocular depth estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3916–3925 (2022)

    2022

  68. [68]

    IEEE Transactions on signal processing40(10), 2570–2583 (2002)

    Zhang, J.: The mean field theory in em procedures for markov random fields. IEEE Transactions on signal processing40(10), 2570–2583 (2002)

    2002

  69. [69]

    In: Robotics: Science and Systems

    Zhang, Z., Suleiman, A.A., Carlone, L., Sze, V., Karaman, S.: Visual-inertial odom- etry on chip: An algorithm-and-hardware co-design approach. In: Robotics: Science and Systems. vol. 13. Robotics: Science and Systems (2017)

    2017

  70. [70]

    In: 2024 International Conference on 3D Vision (3DV)

    Zhu, Z., Peng, S., Larsson, V., Cui, Z., Oswald, M.R., Geiger, A., Pollefeys, M.: Nicer-slam: Neural implicit scene encoding for rgb slam. In: 2024 International Conference on 3D Vision (3DV). pp. 42–52. IEEE (2024)

    2024

  71. [71]

    In: IEEE Conf

    Zhu, Z., Peng, S., Larsson, V., Xu, W., Bao, H., Cui, Z., Oswald, M.R., Pollefeys, M.: Nice-slam: Neural implicit scalable encoding for slam. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12786–12796 (2022)

    2022

  72. [72]

    GT geodesic distance

    Zhuang, B., Cheong, L.F., Lee, G.H.: Baseline desensitizing in translation averag- ing. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4539–4547 (2018) PixVOD 19 A Qualitative Results In the second part of the accompanying videopixvod.mp4, we demonstrate how the proposed method accumulates a trajectory in the Replica scene. Starting from the first frame a...

    work page arXiv 2018