Pith. sign in

REVIEW 4 major objections 6 minor 43 references

Robust 6-DoF Object Pose Tracking with Built-In Recovery under Occlusions and Rapid Object Motions

T0 review · 4 major / 6 minor · reviewed 2026-07-30 · grok-4.5

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

Pith's one-line read A model-based RGB-D tracker can detect its own failures under occlusion and fast motion and automatically re-lock onto the object.

desk verdict Solid ICG+ systems extension with a real detect-and-recover loop; public-benchmark numbers are fine, but the headline robustness win rests on a small, unreleased hard set. read the letter →

arxiv 2607.23468 v1 pith:PIHOGDID submitted 2026-07-26 cs.CV

Balázs Opra , Léo Ghafari , Thomas Stewart , Cyrill Stachniss This is my paper
classification cs.CV
keywords 6-DoFposetrackingRGB-Dfailurerecoveryocclusionkeypointmatchingmodel-basedkeyframestorereal-timerobotics
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

Most real-time 6-DoF object trackers break when the object is fully hidden or moves too fast, and once lost they usually need a human to restart them. This paper shows that a single RGB-D system can both keep accurate track on ordinary sequences and notice when its own pose estimate has diverged, then recover without external supervision. It does so by matching learned keypoints against a stored set of past views, checking whether those matches are still consistent in 3D, and—if not—running a global crop-and-register search that must itself pass the same consistency gate before tracking resumes. On standard benchmarks the method matches leading accuracy at 57.6 frames per second; on a new hard set of occluded and fast-moving scenes, turning recovery on lifts overall accuracy by more than thirty ADD points and beats the compared baselines. A sympathetic reader cares because temporary loss of sight is routine in real robot work, and a tracker that can re-acquire on its own removes a major practical bottleneck.

What carries the argument

The cycle-consistency reliability check: after pose update, matched keypoints are re-projected onto the mesh; a match is an inlier only if the current and keyframe 3D model points agree within a threshold. Enough inliers keep tracking; otherwise the system enters recovery (multi-crop SuperPoint matching against diverse stored keyframes, TEASER++ registration, one refinement iteration, then the same check).

What would settle it

On sequences with full occlusion or sudden motion, disable the recovery module (or empty the keyframe store) and measure whether ADD/ADD-S collapses relative to the full system, and whether recovery still succeeds when the object reappears only from viewpoints never stored as keyframes.

Watch

Extended reading notes

Core claim

Combining efficient SuperPoint matching, a 3D point-to-point texture residual, an icosahedron-indexed keyframe store, and a cycle-consistency reliability check yields a model-based RGB-D tracker that matches state-of-the-art accuracy on ordinary sequences, runs at 57.6 FPS, and—when the consistency check fails—can re-detect the object by multi-crop matching and TEASER registration, then resume only after the recovered pose itself passes the same check, producing the strongest results among the evaluated methods under full occlusion and rapid motion.

Load-bearing premise

Recovery only works if the system has already stored correctly posed views from enough directions; if early poses were wrong or the object returns from an unseen angle, the check and re-detection can fail or lock onto the wrong instance.

Editorial extensions

If this is right

  • Robot manipulation pipelines can keep continuous 6-DoF tracks through temporary full occlusions and drops without manual re-initialization.
  • Built-in self-monitoring of pose reliability can replace external segmentation supervisors for failure detection.
  • Keeping a viewpoint-diverse keyframe store improves both fast-motion tracking and post-failure re-detection compared with a short recent-keyframe queue alone.
  • A texture-plus-depth 3D residual plus denser learned keypoints raises rotational accuracy on standard robotics pose-tracking sets while staying real-time.
  • Recovery latency on the order of a few hundred milliseconds is short enough that freezing the last pose briefly still yields usable overall accuracy on hard sequences.

Reading between the lines

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

  • Pre-filling the keyframe store with rendered views from a textured CAD model before deployment would remove the cold-start view-coverage limit the authors flag.
  • The same cycle-consistency gate could serve as a cheap online confidence signal for downstream grasp or contact planners, not only as a binary recovery trigger.
  • Identity switches among identical instances suggest pairing the recovery stage with a lightweight instance ID or spatial prior when several copies of the same object are present.
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

4 major / 6 minor

Summary. The paper presents a model-based RGB-D 6-DoF object tracker built on ICG+ [30], adding (a) a higher-throughput SuperPoint/TensorRT texture modality with a new 3D point-distance residual and analytic Gauss–Newton terms, (b) an icosahedron-indexed persistent keyframe store, (c) a cycle-consistency reliability check (Eq. 4) that gates keyframe creation and triggers recovery, and (d) a recovery module that matches full-image crops against farthest-point-sampled keyframes and solves TEASER++ registrations per crop, accepting a candidate only after refinement and a repeated reliability check. The authors claim (i) reliable detection of and recovery from tracking failure under full occlusion and rapid motion, (ii) accuracy matching state of the art on standard benchmarks, and (iii) 57.6 FPS tracking with ~210 ms recovery. Evaluation covers YCB-Video, YCBInEOAT, Fast-YCB, and a new in-house dataset of 6 sequences / 2491 frames with fast-motion and occlusion splits, with ablations of the texture stack (Table IV) and a recovery on/off comparison (Table I).

Significance. If the results hold, this is a useful and practically relevant contribution: a self-contained failure-detection-and-recovery loop for model-based tracking at near-60 FPS is a real gap — the paper correctly notes that RGBTrack [6] is essentially the only prior system addressing tracking reliability, and that one depends on an external segmentation module. The manuscript ships several things referees should credit: a recovery on/off ablation with a large measured effect (Table I), a component-wise ablation isolating the 3D residual, SuperPoint, keyframe count, and keyframe store (Table IV), a negative result on LightGlue with concrete runtime numbers (4 ms → 43 ms), competitive accuracy on three public benchmarks against four baselines, and an unusually honest Limitations section (dependence on observed view directions; identity-switch risk with duplicate instances). The system-level design — cycle-consistency gating of both keyframe creation and recovery acceptance — is simple, reproducible, and falsifiable.

major comments (4)
  1. [Sec. IV-B, Table I] Sec. IV-B / Table I: the headline robustness claim (claim (i), 'reliably detect and recover') rests on seven recovery events across six sequences of an unreleased in-house dataset. No per-event statistics are reported: no detection latency distribution, no false-trigger rate of the reliability gate (N_inliers = 10, ε_cc), no recovery success rate conditioned on a true failure, and no precision/recall of the gate against ground-truth visibility. A +32.77 ADD AUC gap over the no-recovery ablation is consistent with anything from 7/7 clean recoveries to a small number of lucky ones. Per-sequence and per-event breakdowns (frames in recovery, time-to-recover, success/failure per event) are needed for claim (i) to be load-bearing; the aggregate AUC alone cannot carry the word 'reliably'.
  2. [Sec. IV-B] Sec. IV-B: the evaluation freezes the output pose for 6 frames after each recovery 'to account for recovery latency.' This is a synthetic post-hoc correction applied only to the authors' method in an otherwise offline evaluation. Its effect on ADD/ADD-S is asymmetric and unquantified: a frozen pose that happens to sit near ground truth inflates the score, one that does not deflates it, and no baseline receives an analogous adjustment. Please either run the system online at the stated 30 FPS and report those numbers, or report Table I both with and without the freeze (and ideally with a sensitivity sweep over freeze length) so the reader can see the correction's contribution.
  3. [Sec. IV-A] Sec. IV-A: ground truth for the in-house dataset — the dataset that carries the central claim — is 'primarily by hand' but 'a subset of frames [is] initialize[d] with FoundationPose,' which is itself a compared baseline in Table I (FP column). The paper does not state which frames, what fraction, or whether/how the initialized frames were manually corrected and verified against the second stereo pair. Residual bias toward FoundationPose-style poses would contaminate the comparison in both directions. At minimum, quantify the subset and describe the verification procedure; ideally, report Table I restricted to fully hand-annotated frames.
  4. [Sec. III-B.2, Eq. (4)] Sec. III-B.2 / Eq. (4): ε_cc is the core threshold of the reliability check — it gates both keyframe creation and recovery acceptance — yet its value is never reported (N_inliers = 10 is). Given that the check's behavior under partial occlusion and symmetric objects (the mug is the weakest row in Table I) depends directly on ε_cc, the paper should state the value used in all experiments and include at least a small sensitivity analysis, or an argument that performance is flat over a wide range.
minor comments (6)
  1. [Secs. III-B/III-C, Table IV] Sec. III-B.1 vs. Sec. III-C and Table IV: keyframe-count parameters are used inconsistently across the text — N_recovery = 20 (Sec. III-C), keyframe store with N_KFS = 10 in Table IV, queue N_KFQ = 9. A single parameter table with all thresholds (ε_cc, ε_3D = 2c, c, σ², λ_R, λ_t, N_KFQ, N_KFS, N_recovery, N_crops, crop scale 0.25, 10°/30°/5° angles) would substantially improve reproducibility.
  2. [Table II, Sec. IV-C] Table II: the claim of 'matching state-of-the-art accuracy' is fair on average (95.0 vs. 96.0 ADD for FoundationPose on YCB-Video; 98.87 vs. 99.05 on Fast-YCB), but the abstract's phrasing 'matches state-of-the-art accuracy on easy tracking sequences' slightly overstates per-object results — e.g., pudding box (76.3 vs. 97.8 for FP) and bowl (84.5 vs. 95.2) are large regressions against ICG+ as well. A sentence explaining these failure modes (textureless/region-dominated objects) would calibrate the claim.
  3. [Sec. IV-A] Sec. IV-A: the dataset URL is embedded mid-sentence in the running text ('Dataset page: github.com/...'); move to a footnote or the conclusion. Also state the capture frame rate of the in-house dataset — the 6-frame freeze is justified by latency 'measured online at 30 FPS,' but the dataset frame rate is not given, so the reader cannot check the correspondence.
  4. [Sec. III-B.3, Eqs. (10)–(11)] Eq. (10)–(11): the sign conventions for g and H (with the (−H + Λ)^{-1} g update in Eq. (3)) are correct but easy to misread; one sentence stating that Eq. (11) is negative semi-definite by construction, hence −H + Λ is positive definite, would help. Also, the ablation comparing the two linearizations (depth-point vs. model-point) reports only YCBInEOAT aggregate numbers (0.49/0.07 points); consider moving it to the ablation section.
  5. [Figs. 2–4] Fig. 3: the inlier/outlier coloring is described in the caption, but the figure would benefit from axis labels or a scale bar indicating what ε_cc corresponds to physically; Fig. 2(e)'s icosahedron is mentioned before it is visually identifiable. Minor readability issues in Fig. 4's step numbering (steps 3 and 1 appear visually out of order in the layout).
  6. [Sec. II / Sec. III-C] Related work: given the recovery-by-relocalization framing, a brief comparison discussion against running a detector-free pose estimator per frame (e.g., FoundPose [21] or MegaPose-style methods) as an alternative recovery backend would strengthen Sec. II; the CNOS+FoundationPose latency comparison in Sec. III-C partially does this but only on runtime, not accuracy.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical systems paper whose accuracy and recovery claims are measured on external benchmarks and ablations, not forced by definition or self-citation.

full rationale

This is a model-based RGB-D tracking systems paper. Its load-bearing claims are empirical (ADD/ADD-S AUC on YCB-Video, YCBInEOAT, Fast-YCB, and an in-house set; FPS; recovery ablation in Table I), not first-principles derivations. The optimization (Gauss-Newton on region/depth/texture PDFs), SuperPoint matching, 3D distance residual, icosahedron keyframe store, cycle-consistency gate (N_inliers, ε_cc), and TEASER++ recovery are design choices evaluated against held-out sequences and a recovery-disabled ablation (+32.77 ADD AUC when recovery is on). Thresholds and keyframe counts are tuning knobs, not quantities fitted to the reported metric and then re-presented as predictions. Citations (ICG+, SuperPoint, TEASER++, FoundationPose, etc.) are external prior art, not author-overlapping uniqueness theorems that force the result. Evaluation-quality concerns (few recovery events, pose-freeze protocol, partial FoundationPose annotation of the unreleased set) affect confidence in the robustness claim but do not make any equation or claim reduce to its inputs by construction. No circular step applies.

Assumptions & free parameters 8 free parameters · 6 assumptions · 3 invented entities

The central robustness claim rests on standard calibrated RGB-D tracking assumptions, the ICG+ multi-modal Gauss-Newton backbone, learned SuperPoint descriptors as a black-box matcher, TEASER++ as a robust registrar, and several hand-set reliability/recovery thresholds. No new physical entities are postulated; invented pieces are engineering modules (reliability check, keyframe store, recovery controller).

free parameters (8)
  • Ninliers (reliability gate) = 10
    Minimum inlier cycle-consistent matches required to accept a pose / exit recovery; stated as found to work well.
  • ϵ_cc cycle-consistency distance = positive scalar, value not numerically fixed in text
    3D threshold deciding inlier vs outlier when comparing keyframe and current lifted keypoints in the model frame.
  • ϵ_3D outlier reject (=2c) and Tukey c = ϵ_3D=2c
    3D match rejection and robust loss scale tied to expected inter-frame surface motion; ϵ_3D=2c used in all experiments.
  • Ncrops and Nrecovery = 16 and 20
    Image crop count and number of farthest-point keyframes used in recovery; dominate recovery latency (~210 ms).
  • Keyframe creation angle and store selection cone = 10° / 30° / 5°; NKFQ=9, NKFS=10 in ablation
    New keyframe if orientation changes >10°; store samples within 30° cone with ≥5° separation; NKFQ and NKFS sizes (ablation uses 9 and 10).
  • SuperPoint ROI and keypoint count = 1024px ROI, 1800 keypoints
    1024×1024 crop with 1800 keypoints chosen for throughput/accuracy tradeoff vs SIFT baseline.
  • Tikhonov λ_R, λ_t and texture variance σ² = not numerically specified
    Regularization and modality weighting in the joint Gauss-Newton solve; inherited/tuned as part of ICG+-style optimization.
  • Post-recovery pose freeze length in evaluation = 6 frames
    Output pose frozen for 6 frames after recovery to account for average recovery latency at 30 FPS when scoring the hard dataset.
assumptions (6)
  • domain assumption A CAD model or reconstructed mesh of the target is available at test time, with known camera intrinsics and extrinsics (WTC).
    Stated at the opening of Sec. III; defines the model-based tracking setting.
  • domain assumption Multi-modal correspondences (region, depth, texture) can be fused in a Gauss-Newton se(3) update as in ICG+ with first-order exponential-map pose increments.
    Sec. III-A restates the ICG+ posterior product and Tikhonov-regularized solve as the tracking backbone.
  • domain assumption SuperPoint descriptors plus nearest-neighbor matching with Lowe ratio and two-view cycle checks yield usable 2D-3D object correspondences under the paper's crops.
    Sec. III-B; LightGlue was tried and rejected on speed, not accuracy.
  • ad hoc to paper If keyframe pose, current pose, and matches are all correct, lifted matched keypoints coincide in the model frame within ϵ_cc (cycle-consistency reliability).
    Sec. III-B.2 Eq. (4); this is the paper's failure detector, not a standard external theorem.
  • domain assumption TEASER++ on depth-lifted crop-to-keyframe matches produces a coarse pose good enough that one tracking iteration plus the reliability check can accept or reject recovery.
    Sec. III-C recovery pipeline; relies on TEASER++'s robust registration guarantees under outlier correspondences.
  • domain assumption ADD and ADD-S AUC are appropriate proxies for tracking quality in robotics manipulation settings.
    Used throughout Sec. IV following Hinterstoisser et al. / YCB-Video practice.
invented entities (3)
  • Cycle-consistency tracking reliability check (inlier count over keyframes)
    purpose: Detect divergence/occlusion without an external segmentation supervisor and gate keyframe creation.
    Core novel control signal of the system (Sec. III-B.2); evaluated indirectly via recovery ablations rather than calibrated detector metrics.
  • Icosahedron-indexed persistent keyframe store
    purpose: Retain viewpoint-diverse object appearance for fast motion and post-failure re-detection beyond a short keyframe queue.
    Sec. III-B.1; engineering structure built on subdivided icosahedron view bins.
  • Multi-crop TEASER recovery controller with post-refine reliability gate
    purpose: Globally re-detect the object after failure and avoid accepting spurious poses.
    Sec. III-C and Fig. 4; composes known matchers/registrars into a recovery mode with stated ~210 ms cost.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust 6-DoF Object Pose Tracking with Built-In Recovery under Occlusions and Rapid Object Motions." pith.science (2026). https://pith.science/paper/PIHOGDID

@misc{pith2026260723468,
  author       = {Pith},
  title        = {Pith review of: Robust 6-DoF Object Pose Tracking with Built-In Recovery under Occlusions and Rapid Object Motions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PIHOGDID}},
  note         = {Machine review of arXiv:2607.23468}
}
read the original abstract

Real-time 6-DoF object pose tracking is essential for many robotics applications, and several approaches exist. Yet even today's approaches remain unreliable under temporary full occlusions and rapid object motions. Once tracking is lost, most methods struggle to detect the failure and recover automatically, often requiring manual re-initialization. In this paper, we address the problem of robust model-based 6-DoF tracking of unseen objects from RGB-D data, especially in scenarios with occlusion and fast motion. We propose a novel method that combines efficient learning-based keypoint matching with optimization-based alignment and introduces a novel failure detection and recovery module. Our system monitors pose reliability, detects tracking divergence or occlusions, and performs a global re-detection and pose estimation step that robustly verifies recovery candidates before resuming tracking. Our evaluation on standard tracking benchmarks and on a new dataset of occluded and fast-moving scenes shows that our method matches state-of-the-art accuracy on easy tracking sequences, maintains high tracking speed at 57.6 frames per second, and provides the most robust tracking performance under challenging conditions. Thus, we believe that our approach is a relevant step forward in robust 6-DoF object tracking from RGB-D data.

Figures

Figures reproduced from arXiv: 2607.23468 by the authors.

Figure 1
Figure 1. Our 6-DoF object pose tracking system is capable of [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the Texture-Based Tracking System. (a) We perform texture-based keypoint matching between frozen prior frames [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Tracking reliability check. The keyframe-current frame [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Failure recovery procedure: 1) we match well-selected keyframe store keyframes with all crops of the current image, 2) these [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 5 linked inside Pith

  1. [30]

    Stoiber, M

    M. Stoiber, M. Elsayed, A.E. Reichert, F. Steidle, D. Lee, and R. Triebel. Fusing Visual Appearance and Geometry for Multi- Modality 6DoF Object Tracking. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS) , 2023

    2023

  2. [6]

    Guo and J

    T. Guo and J. Yu. RGBTrack: Fast, Robust Depth-Free 6D Pose Estimation and Tracking. arXiv preprint, arXiv:2506.17119, 2025

    arXiv 2025

  3. [1]

    Cheng and A.G

    H.K. Cheng and A.G. Schwing. XMem: Long-term video object segmentation with an atkinson-shiffrin memory model. In Proc. of the Europ. Conf. on Computer Vision (ECCV) , page 640–658, 2022

    2022

  4. [2]

    W. Deng, D. Campbell, C. Sun, J. Zhang, S. Kanitkar, M.E. Shaffer, et al. Pos3R: 6D Pose Estimation for Unseen Objects Made Easy. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) , 2025

    2025

  5. [3]

    X. Deng, A. Mousavian, Y . Xiang, F. Xia, T. Bretl, and D. Fox. PoseRBPF: A Rao-Blackwellized Particle Filter for 6D Object Pose Tracking. arXiv preprint, arXiv:1905.09304, 2019

    arXiv 1905

  6. [4]

    DeTone, T

    D. DeTone, T. Malisiewicz, and A. Rabinovich. SuperPoint: Self- Supervised Interest Point Detection and Description. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2017

    2017

  7. [5]

    Ge and G

    R. Ge and G. Loianno. VIPose: Real-time Visual-Inertial 6D Object Pose Tracking. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS) , pages 4597–4603, 2021

    2021

  8. [7]

    Harris and M

    C. Harris and M. Stephens. A combined corner and edge detector. In Alvey Vision Conf. , 1988

    1988

Show all 43 references
  1. [8]

    Hartley and A

    R. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, second edition, 2004

    2004

  2. [9]

    Hinterstoisser, V

    S. Hinterstoisser, V . Lepetit, S. Ilic, S. Holzer, G. Bradski, K. Konolige, et al. Model Based Training, Detection and Pose Estimation of Texture-Less 3D Objects in Heavily Cluttered Scenes. In Proc. of the Asian Conf. on Computer Vision (ACCV) , pages 548–562, 2012

    2012

  3. [10]

    Huang, J

    B. Huang, J. Yu, and S. Jain. EARL: Eye-on-Hand Reinforcement Learner for Dynamic Grasping with Active Pose Estimation. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 2963–2970, 2023

    2023

  4. [11]

    W. Kehl, F. Tombari, S. Ilic, and N. Navab. Real-Time 3D Model Tracking in Color and Depth on a Single CPU Core. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2017

    2017

  5. [12]

    Klein and D

    G. Klein and D. Murray. Parallel Tracking and Mapping for Small AR Workspaces. In Proc. of the Intl. Symp. on Mixed and Augmented Reality (ISMAR) , 2007

    2007

  6. [13]

    Leroy, Y

    V . Leroy, Y . Cabon, and J. Revaud. Grounding Image Matching in 3D with MASt3R. In Proc. of the Europ. Conf. on Computer Vision (ECCV), pages 71–91, 2024

    2024

  7. [14]

    Y . Li, G. Wang, X. Ji, Y . Xiang, and D. Fox. DeepIM: Deep Iterative Matching for 6D Pose Estimation. arXiv preprint, arXiv:1804.00175, 2018

    2018 arXiv

  8. [15]

    Liang, Y

    T. Liang, Y . Zeng, J. Xie, and B. Zhou. DynamicPose: Real-time and Robust 6D Object Pose Tracking for Fast-Moving Cameras and Objects. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS) , pages 2424–2431, 2025

    2025

  9. [16]

    Lindenberger, P.E

    P. Lindenberger, P.E. Sarlin, and M. Pollefeys. LightGlue: Local fea- ture matching at light speed. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV) , pages 17581–17592, 2023

    2023

  10. [17]

    D. Lowe. Distinctive Image Features from Scale-Invariant Keypoints. Intl. Journal of Computer Vision (IJCV) , 60(2):91–110, 2004

    2004

  11. [18]

    Nguyen, C

    V .N. Nguyen, C. Forster, S. Shkodrani, V . Lepetit, B. Tekin, C. Keskin, et al. GoTrack: Generic 6DoF Object Pose Refinement and Tracking. arXiv preprint, arXiv:2506.07155, 2025

    2025 arXiv

  12. [19]

    Nguyen, T

    V .N. Nguyen, T. Groueix, G. Ponimatkin, V . Lepetit, and T. Hodan. CNOS: A Strong Baseline for CAD-based Novel Object Segmentation. In Proc. of the Intl. Conf. on Computer Vision Workshops , 2023

    2023

  13. [20]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H.V . V o, M. Szafraniec, V . Khalidov, et al. DINOv2: Learning Robust Visual Features without Supervision. Trans. on Machine Learning Research (TMLR) , 2024

    2024

  14. [21]

    Ornek, Y

    E.P. Ornek, Y . Labbe, B. Tekin, L. Ma, C. Keskin, C. Forster, et al. FoundPose: Unseen Object Pose Estimation with Foundation Features. In Proc. of the Europ. Conf. on Computer Vision (ECCV) , pages 163– 182, 2024

    2024

  15. [22]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, et al. PyTorch: An Imperative Style, High-Performance Deep Learning Library. In Proc. of the Conf. on Neural Information Processing Systems (NeurIPS) , 2019

    2019

  16. [23]

    N.A. Piga, Y . Onyshchuk, G. Pasquale, U. Pattacini, and L. Natale. ROFT: Real-Time Optical Flow-Aided 6D Object Pose and Velocity Tracking. IEEE Robotics and Automation Letters (RA-L) , 7(1):159– 166, 2022

    2022

  17. [24]

    Prisacariu and I.D

    V .A. Prisacariu and I.D. Reid. PWP3D: Real-Time Segmentation and Tracking of 3D Objects. Intl. Journal of Computer Vision (IJCV) , 98(3):335–354, 2012

    2012

  18. [25]

    K. Rana, J. Abou-Chakra, S. Garg, R. Lee, I. Reid, and N. S ¨underhauf. Affordance-Centric Policy Learning: Sample Efficient and Generalis- able Robot Policy Learning using Affordance-Centric Task Frames. arXiv preprint, arXiv:2410.12124, 2024

    2024 arXiv

  19. [26]

    Ren, V .A

    C.Y . Ren, V .A. Prisacariu, O. K ¨ahler, I.D. Reid, and D.W. Murray. Real-Time Tracking of Single and Multiple Objects from Depth- Colour Imagery Using 3D Signed Distance Functions. Intl. Journal of Computer Vision (IJCV) , 124(1):80–95, 2017

    2017

  20. [27]

    Shaikewitz, S

    L. Shaikewitz, S. Ubellacker, and L. Carlone. A certifiable algorithm for simultaneous shape estimation and object tracking. IEEE Robotics and Automation Letters (RA-L) , 9(12):11873–11880, 2024

    2024

  21. [28]

    Shankar, M

    K. Shankar, M. Tjersland, J. Ma, K. Stone, and M. Bajracharya. A Learned Stereo Depth System for Robotic Manipulation in Homes. IEEE Robotics and Automation Letters (RA-L) , 7(2):2305–2312, 2022

    2022

  22. [29]

    Shi and C

    J. Shi and C. Tomasi. Good Features to Track. Technical report, Cornell University, 1993

    1993

  23. [31]

    Stoiber, M

    M. Stoiber, M. Pfanne, K.H. Strobl, R. Triebel, and A. Albu-Sch ¨affer. SRT3D: A Sparse Region-Based 3D Object Tracking Approach for the Real World. Intl. Journal of Computer Vision (IJCV) , 130(4):1008– 1030, 2022

    2022

  24. [32]

    Stoiber, M

    M. Stoiber, M. Sundermeyer, and R. Triebel. Iterative Corresponding Geometry: Fusing Region and Depth for Highly Efficient 3D Tracking of Textureless Objects. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) , 2022

    2022

  25. [33]

    Tjaden, U

    H. Tjaden, U. Schwanecke, and E. Sch ¨omer. Real-Time Monocular Pose Estimation of 3D Objects using Temporally Consistent Local Color Histograms. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2017

    2017

  26. [34]

    Vacchetti, V

    L. Vacchetti, V . Lepetit, and P. Fua. Stable Real-Time 3D Tracking Using Online and Offline Information. IEEE Trans. on Pattern Analysis and Machine Intelligence (TPAMI) , 26(10):1385–1391, 2004

    2004

  27. [35]

    C. Wang, R. Mart ´ın-Mart´ın, D. Xu, J. Lv, C. Lu, F.F. Li, et al. 6- PACK: Category-level 6D Pose Tracker with Anchor-Based Keypoints. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA) , pages 10059–10066, 2020

    2020

  28. [36]

    Wen and K

    B. Wen and K. Bekris. BundleTrack: 6D Pose Tracking for Novel Objects without Instance or Category-Level 3D Models. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS) , pages 8067–8074, 2021

    2021

  29. [37]

    B. Wen, C. Mitash, B. Ren, and K.E. Bekris. se(3)-TrackNet: Data- driven 6D Pose Tracking by Calibrating Image Residuals in Synthetic Domains. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS) , 2020

    2020

  30. [38]

    B. Wen, J. Tremblay, V . Blukis, S. Tyree, T. M ¨uller, A. Evans, et al. BundleSDF: Neural 6-DoF Tracking and 3D Reconstruction of Unknown Objects. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) , 2023

    2023

  31. [39]

    B. Wen, W. Yang, J. Kautz, and S. Birchfield. FoundationPose: Unified 6D Pose Estimation and Tracking of Novel Objects. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 17868–17879, 2024

    2024

  32. [40]

    Xiang, T

    Y . Xiang, T. Schmidt, V . Narayanan, and D. Fox. PoseCNN: A Convolutional Neural Network for 6D Object Pose Estimation in Cluttered Scenes. In Proc. of Robotics: Science and Systems (RSS) , 2018

    2018

  33. [41]

    H. Yang, J. Shi, and L. Carlone. TEASER: Fast and Certifiable Point Cloud Registration. IEEE Trans. on Robotics (TRO) , 37(2):314–333, 2020

    2020

  34. [42]

    Zhong and L

    L. Zhong and L. Zhang. A Robust Monocular 3D Object Tracking Method Combining Statistical and Photometric Constraints. Intl. Jour- nal of Computer Vision (IJCV) , 127(8):973–992, 2019

    2019

  35. [43]

    Zhou and K

    Y . Zhou and K. Yang. Exploring TensorRT to Improve Real-Time Inference for Deep Learning. In Proc. of the Intl. Conf. on High Performance Computing & Communications , pages 2011–2018, 2022

    2011

Pith tools

Reviewed July 30, 2026 · model on record in the stance chip above.