Pith. sign in

REVIEW 4 major objections 27 references

Indoor robots can re-identify generic objects after long absences from monocular RGB alone, without pose or depth, by accumulating appearance memory and spatial context.

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 ↗

REMIND combines frozen DINOv3 features, dual-bank multi-prototype memory, part/background descriptors, neighbour co-occurrence, and joint Hungarian assignment to re-identify generic indoor objects after long gaps, reaching 90.35% IDF1 on a custom revisit sequence.

T0 review reviewed 2026-07-13 challenge →

load-bearing objection Solid systems paper for monocular long-gap indoor re-ID; the big custom-video margin is real but load-bearing on static layout and neighbour context, while ScanNet++ gains are smaller and partly inverted under YOLO. the 4 major comments →

arxiv 2607.09267 v1 pith:WSQQ7E62 submitted 2026-07-10 cs.CV

REMIND: RE-Identification with Memory for INDoor Navigation

Pablo Diaz-Pereda , Alejandro Rodriguez-Ramos , David Perez-Saura , Pascual Campoy This is my paper
classification cs.CV
keywords long-term re-identificationmulti-object trackingindoor navigationappearance memoryneighbour contextmonocular RGBDINOv3identity persistence
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.

The reading

Mobile robots that leave a room and return later must re-identify the same chairs, backpacks, monitors and other everyday objects under new viewpoints, lighting and partial occlusion. Existing multi-object trackers are built for short gaps between video frames of people or cars; person re-identification lacks lasting memory; and video-object-segmentation methods filter distractors reactively instead of enforcing global identity consistency. REMIND treats the problem as long-term multi-object re-identification of generic indoor objects from monocular RGB only. Motivated by human visual cognition that relies on accumulated appearance familiarity and co-occurring spatial context rather than self-localisation, it freezes a foundation-model feature extractor, stores multi-prototype appearance banks (global, part and background), reasons over neighbour co-occurrence, and solves joint assignment with ambiguity safeguards. On a purpose-built re-entry sequence with dense same-class clutter it reaches 90.35 % IDF1—nearly 20 points above a strong video-object-segmentation baseline and more than 36 above a tracking-by-detection baseline—while completing every ScanNet++ scene where the segmentation baseline exhausts GPU memory on two-thirds of the detection runs.

Core claim

Long-term multi-object re-identification of generic indoor objects from monocular RGB, without camera pose or depth, is solved to high accuracy by combining frozen DINOv3 features, a dual-bank multi-prototype appearance memory (global, part and background descriptors), neighbour-context reasoning from spatial co-occurrence, and joint Hungarian assignment with ambiguity-aware guards. On a controlled re-entry sequence the system attains 90.35 % IDF1, nearly 20 points above DAM4SAM and more than 36 above MASA under ground-truth masks, and leads IDF1 in almost every ScanNet++ regime while finishing 100 % of scenes.

What carries the argument

REMIND’s dual-bank multi-prototype memory (work bank for recent observations, stable bank for consolidated prototypes) together with part- and background-level descriptors, a neighbour-context graph that supplies co-occurrence bonuses or vetoes, and global per-class Hungarian assignment with STRONG/AMBIGUOUS/WEAK guards.

Load-bearing premise

Objects stay put while the camera moves, so historical co-occurrence and pairwise spatial-distance graphs remain trustworthy identity cues; if furniture is rearranged between visits those cues become misleading.

What would settle it

Run the same controlled re-entry sequence after deliberately swapping or relocating several same-class objects between the robot’s exit and re-entry; if IDF1 and recovery collapse relative to the static-scene baseline, the neighbour-context premise is falsified.

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

If this is right

  • Robots can maintain persistent object identities across room exits and re-entries using only monocular RGB, removing the need for continuous pose estimation or depth sensing for this task.
  • Appearance memory plus spatial co-occurrence is a sufficient computational primitive for long-gap indoor re-identification, reducing dependence on 3-D reconstruction pipelines.
  • Global joint assignment with ambiguity safeguards yields higher recovery and association accuracy than independent per-object memory banks used by current video-object-segmentation methods.
  • The released custom re-entry dataset and evaluation protocol become the first dedicated baseline for measuring long-term indoor identity persistence under controlled revisits and same-class clutter.

Where Pith is reading between the lines

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

  • The same dual-bank and neighbour-context design could be tested on outdoor mobile platforms where objects are mostly static (parked cars, street furniture) but viewpoint and lighting change dramatically.
  • Tight coupling of the detector and the memory banks—feeding re-identification confidence back into mask selection—would likely close the remaining gap between ground-truth-mask and end-to-end YOLO performance.
  • If objects can move, the system would need an explicit motion model or an online graph-update rule that forgets outdated co-occurrence edges, turning the current static-scene premise into a controllable module.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 0 minor

Summary. REMIND is an online multi-object re-identification tracker for generic indoor objects from monocular RGB, without camera pose or depth. It combines frozen DINOv3 patch features with dual-bank multi-prototype memory (global, part, and background channels), neighbour co-occurrence and spatial-distance context, and class-wise Hungarian assignment with ambiguity-aware guards. On a purpose-built 2 min 17 s indoor re-entry sequence with dense same-class clutter, it reports 90.35% IDF1 under ground-truth masks (≈20 points above DAM4SAM, >36 above MASA) and retains a lead under YOLO; on ScanNet++ it leads IDF1 in most regimes, completes all scenes, and shows higher AssA/recovery than MASA even where MASA edges all-scenes YOLO IDF1. Ablations attribute large gains to background rings and neighbour context; system, evaluation code, and the custom sequence are released.

Significance. Long-term identity persistence for generic indoor objects under robot re-entry is a practically important and underexplored setting at the MOT/Re-ID/VOS intersection. The paper’s contribution is a carefully engineered, fully automatic monocular pipeline with joint assignment and persistent memory, plus recovery-oriented metrics, behaviour breakdowns, runtime/memory profiles, and a public release of code, evaluation framework, and a controlled re-entry sequence. If the reported margins hold under broader revisit conditions, REMIND would be a useful deployable baseline for indoor navigation without depth or pose. Credit is due for dual evaluation regimes (GT masks vs YOLO), explicit ablations (background, neighbours, greedy assignment), and honest reporting of ScanNet++ trade-offs (ID-switch rates, DetA gaps, MASA’s all-scenes YOLO edge).

major comments (4)
  1. The abstract and §IV-B headline (90.35% IDF1, ~20 pts over DAM4SAM, >36 over MASA under GT masks) rest on a single author-designed monocular sequence of 2 min 17 s (§IV-A). Table VI shows that removing neighbour-context alone drops custom IDF1 from 90.35% to 69.73% (−20.6 pts) while costing only ~2.1 pts on ScanNet++; removing background collapses custom IDF1 to 55.56%. The largest claimed margin is therefore load-bearing on this one static re-entry trajectory and on the neighbour/background modules, not on a multi-sequence distribution of revisits. Either expand the custom benchmark (multiple rooms/devices/revisit patterns, as promised in future work) or substantially re-scope the abstract and conclusions so the 20-point claim is clearly conditioned on that single controlled sequence rather than presented as the primary evidence of general long-term re-ID superiority.
  2. §II-A states that “objects are static, the camera moves,” and §III-D builds co-occurrence and pairwise spatial-distance graphs used for contextual bonuses/penalties and relational disambiguation. Under rearrangement between visits those cues become misleading rather than corrective, yet no controlled stress test (e.g., moved chairs/backpacks between re-entries) is reported. Given that Table VI attributes a large fraction of the custom-video gain to the neighbour layer, the manuscript should (i) state the static-layout premise as an explicit limitation of the current evaluation, and (ii) either add a rearrangement experiment or quantify how often neighbour evidence is decisive versus appearance-only on the custom sequence. Without this, the central claim that “appearance memory and contextual reasoning … are sufficient” is only demonstrated under unchanging layout.
  3. §IV-B acknowledges that the YOLO regime stresses DAM4SAM beyond its one-shot design and that the 66.9% OOM rate is partly caused by instantiating a new SAM2 bank per unmatched detection. The abstract and conclusions still present “completes every scene vs 66.9% OOM” as a primary system advantage without that qualification. Please rephrase the abstract/results summary so the OOM comparison is clearly attributed to architectural mismatch under the forced multi-detection protocol, and keep the GT-mask regime (where DAM4SAM completes almost all scenes) as the primary identity-quality comparison. Also report peak track/memory counts under YOLO so readers can separate algorithmic recovery from VRAM scaling.
  4. Appendix Table VIII lists many free thresholds (τ_match, channel weights, dual-bank τ_dup/α/merge, lifecycle h_confirm/m_max/TTL, δ±, τ_lock, part/background geometry). Only three coarse ablations are given (greedy, no background, no neighbours). Because τ_match is described as “most critical,” a short sensitivity sweep on τ_match and on the background/neighbour caps (δ+, δ−) on both benchmarks would strengthen the claim that gains are architectural rather than over-tuned to the custom sequence. Without it, reproducibility of the headline margins is harder to assess despite the code release.

Circularity Check

0 steps flagged

No circularity: empirical systems paper with external tracking metrics, independent ablations, and no self-definitional or fitted-as-prediction chain.

full rationale

REMIND is an engineering/systems contribution, not a first-principles derivation. The central claims are measured IDF1, AssA, recovery, and hard-recovery on held-out sequences (custom video and ScanNet++) against external baselines (DAM4SAM, MASA) under GT-mask and YOLO regimes. Those scores are standard MOT/HOTA-style quantities computed from predicted vs. ground-truth identity timelines; they are not defined by, or algebraically forced by, the dual-bank memory, neighbour graphs, or Hungarian costs. Ablations (Tables VI–VII) remove modules and report large independent drops (e.g., no-background 90.35%→55.56% IDF1 on the custom video; no-neighbors 90.35%→69.73%), which is falsifiable evidence rather than restatement of inputs. Thresholds (Table VIII) are free heuristic parameters, not parameters fitted to a subset and then re-reported as predictions of a closely related quantity. Motivation cites external visual-cognition work (Oliva & Torralba) and external foundation models (DINOv3, SAM2 ecosystem); there is no load-bearing uniqueness theorem or ansatz imported from overlapping-author prior work. Author-designed custom sequence and the static-scene premise affect generalization risk, not circularity of the derivation chain. Score 0.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 3 invented entities

REMIND is a heuristic online tracker. The central empirical claim rests on frozen foundation features, an off-the-shelf class-aware detector, a large set of hand-chosen thresholds, and domain assumptions that indoor objects are static and that monocular appearance plus co-occurrence suffice without pose or depth. No new physical entities are postulated; invented entities are architectural modules. Free parameters dominate the ledger because association quality is threshold-gated rather than learned end-to-end.

free parameters (7)
  • τ_match (minimum similarity for assignment)
    Called the most critical association parameter in Table VIII; directly gates whether a detection is linked to an existing track or starts a new identity.
  • Channel weights w_obj, w_bg, w_parts and quality floors
    Nominal multi-channel combination weights adjusted by per-detection quality factors; control how much background/parts can rescue weak global appearance.
  • Dual-bank thresholds τ_dup, α, τ_merge,internal, promotion hit counts
    Control prototype insert vs EMA, merging, and work→stable promotion; shape long-term appearance models without learning.
  • Lifecycle thresholds h_confirm, m_max, TTL
    Determine when tracks become CONFIRMED/INACTIVE and when ambiguous or inactive tracks are purged; affect recovery vs switch trade-offs.
  • Context caps δ+, δ− and anchor margin δ_confirm / τ_lock
    Bound neighbour-context bonuses/penalties and pre-lock high-confidence pairs before Hungarian assignment.
  • Part/background geometry (K, ρ, r_in, r_out, τ_merge)
    Hand-chosen descriptor construction parameters for K-means/attention parts and concentric background rings.
  • MASA embedding momentum 0.3 (baseline adaptation)
    Selected as best among tested values for the adapted MASA baseline; affects comparative claims.
axioms (6)
  • domain assumption Frozen DINOv3 (or compatible DINO-family) patch features are sufficiently discriminative for generic indoor object re-identification without task-specific fine-tuning.
    Section III-B adopts a single frozen ViT-S/16 forward pass as the sole visual backbone for all descriptor channels.
  • domain assumption Indoor objects of interest are static while the camera moves; long gaps are due to robot motion, not object motion.
    Stated when contrasting MOT assumptions (Section II-A); co-occurrence and distance graphs accumulate under this premise (III-D).
  • domain assumption Association may be restricted to the same semantic class from an off-the-shelf detector vocabulary.
    System overview and association stage require class-matched candidates; not zero-shot across classes.
  • ad hoc to paper Appearance familiarity plus spatial co-occurrence context, without explicit self-localization or depth, is sufficient for robust long-term object re-identification.
    Motivating principle from visual cognition (Introduction) and restated in Conclusions as the design thesis being tested.
  • ad hoc to paper Hungarian bipartite matching on a multi-channel cost matrix with ambiguity guards yields globally consistent identities better than independent per-object VOS memories.
    Core association design (III-E) and main contrast to DAM4SAM's per-object banks (II-D).
  • standard math Standard bipartite matching / cosine similarity / EMA update mathematics.
    Used throughout association and memory update without novel theorems.
invented entities (3)
  • Dual-bank multi-prototype appearance memory (work + stable banks per descriptor channel) no independent evidence
    purpose: Accumulate long-term appearance while separating recent variation from consolidated prototypes for re-ID after long absences.
    Architectural construct introduced in III-C; not an external physical entity. Independent evidence is only via ablations/metrics in this paper.
  • Neighbour-sets context layer with co-occurrence and spatial-distance graphs no independent evidence
    purpose: Disambiguate same-class candidates and rescue weak appearance matches using historical spatial context.
    Module in III-D; ablation shows large custom-video drop when removed. No independent external validation beyond this work.
  • AmbiguousTrack / ProvisionalNewTrack post-assignment guards no independent evidence
    purpose: Defer hard decisions under ambiguity to reduce premature ID proliferation and switches.
    Heuristic entities in III-E; diagnostic stats given but no external theory.

reviewed 2026-07-13 · how reviews work

0 comments
Cite this review

Pith. "Pith review of REMIND: RE-Identification with Memory for INDoor Navigation." pith.science (2026). https://pith.science/paper/WSQQ7E62

@misc{pith2026260709267,
  author       = {Pith},
  title        = {Pith review of: REMIND: RE-Identification with Memory for INDoor Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WSQQ7E62}},
  note         = {Machine review of arXiv:2607.09267}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Mobile robots operating indoors must re-identify previously observed objects after long temporal gaps, significant viewpoint changes, and severe illumination variations. This remains a challenging problem: multi-object tracking methods are optimized for short-term association of pedestrians and vehicles at video rates, person and vehicle re-identification approaches lack persistent memory mechanisms, and state-of-the-art video object segmentation techniques rely on reactive distractor filtering rather than enforcing global identity consistency. To address these limitations, we present REMIND, an online tracker designed for long-term multi-object re-identification of generic indoor objects from monocular RGB imagery, requiring neither camera pose nor depth. Motivated by evidence from visual cognition that humans rely on accumulated appearance familiarity and spatial context rather than explicit self-localization, REMIND combines frozen DINOv3 features with a dual-bank multi-prototype appearance memory, part- and background-level descriptors, a neighbour-context reasoning module exploiting spatial co-occurrence, and joint Hungarian assignment with ambiguity-aware safeguards. On a purpose-built indoor dataset featuring controlled revisits and dense same-class clutter, REMIND reaches 90.35% IDF1, nearly 20 points above a state-of-the-art video object segmentation baseline and more than 36 above a strong tracking-by-detection baseline. On ScanNet++, it attains the highest IDF1 in every setting but one, end-to-end detection over all scenes, where the tracking-by-detection baseline is marginally ahead while REMIND still associates and recovers identities more accurately; it also completes every scene, whereas the video object segmentation baseline exhausts GPU memory on 66.9% under YOLO detections. The complete system, evaluation framework, and dataset are publicly released.

Figures

Figures reproduced from arXiv: 2607.09267 by Alejandro Rodriguez-Ramos, David Perez-Saura, Pablo Diaz-Pereda, Pascual Campoy.

Figure 1
Figure 1. Figure 1: Long-term object re-identification scenario for indoor [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: End-to-end REMIND pipeline. Perception: a single forward pass of a frozen DINOv3 ViT-S/16 produces patch features shared across all objects in the frame, while a YOLO segmentation detector supplies per-object masks; for each detection, the system extracts complementary global (ϕˆ gl ), part (ϕˆ pt ), and background (ϕˆ bg ) descriptors. Association: a multi￾channel cost matrix is enriched with neighbour-co… view at source ↗
Figure 3
Figure 3. Figure 3: Three-channel descriptor extraction from a single [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Behaviour analysis under different same-class distractor [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Behaviour analysis under different observation mask [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Behaviour analysis under different visible object den [PITH_FULL_IMAGE:figures/full_fig_p010_6.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

27 extracted references · 7 linked inside Pith

  1. [1]

    The role of context in object recognition,

    A. Oliva and A. Torralba, “The role of context in object recognition,” Trends in Cognitive Sciences, vol. 11, no. 12, pp. 520–527, 2007

    2007

  2. [2]

    Simple online and realtime track- ing with a deep association metric,

    N. Wojke, A. Bewley, and D. Paulus, “Simple online and realtime track- ing with a deep association metric,” inIEEE International Conference on Image Processing (ICIP), 2017

    2017

  3. [3]

    ByteTrack: Multi-object tracking by associating every detection box,

    Y . Zhang, P. Sun, Y . Jiang, D. Yu, F. Weng, Z. Yuan, P. Luo, W. Liu, and X. Wang, “ByteTrack: Multi-object tracking by associating every detection box,” inEuropean Conference on Computer Vision (ECCV), 2022

    2022

  4. [4]

    StrongSORT: Make DeepSORT great again,

    Y . Du, Z. Zhao, Y . Song, Y . Zhao, F. Su, T. Gong, and H. Meng, “StrongSORT: Make DeepSORT great again,”IEEE Transactions on Multimedia, 2023

    2023

  5. [5]

    Deep OC-SORT: Multi-pedestrian tracking by adaptive re-identification,

    G. Maggiolino, A. Ahmad, J. Cao, and K. Kitani, “Deep OC-SORT: Multi-pedestrian tracking by adaptive re-identification,”arXiv preprint arXiv:2302.11813, 2023

    Pith/arXiv arXiv 2023

  6. [6]

    MeMOTR: Long-term memory-augmented trans- former for multi-object tracking,

    R. Gao and L. Wang, “MeMOTR: Long-term memory-augmented trans- former for multi-object tracking,” inIEEE International Conference on Computer Vision (ICCV), 2023

    2023

  7. [7]

    Scalable person re-identification: A benchmark,

    L. Zheng, L. Shen, L. Tian, S. Wang, J. Wang, and Q. Tian, “Scalable person re-identification: A benchmark,” inIEEE International Confer- ence on Computer Vision (ICCV), 2015. 12

    2015

  8. [8]

    Person transfer GAN to bridge domain gap for person re-identification,

    L. Wei, S. Zhang, W. Gao, and Q. Tian, “Person transfer GAN to bridge domain gap for person re-identification,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  9. [9]

    Deep relative distance learning: Tell the difference between similar vehicles,

    H. Liu, Y . Tian, Y . Yang, L. Pang, and T. Huang, “Deep relative distance learning: Tell the difference between similar vehicles,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016

    2016

  10. [10]

    TransReID: Transformer-based object re-identification,

    S. He, H. Luo, P. Wang, F. Wang, H. Li, and W. Jiang, “TransReID: Transformer-based object re-identification,” inIEEE International Con- ference on Computer Vision (ICCV), 2021

    2021

  11. [11]

    Bag of tricks and a strong baseline for deep person re-identification,

    H. Luo, Y . Gu, X. Liao, S. Lai, and W. Jiang, “Bag of tricks and a strong baseline for deep person re-identification,” inIEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2019

    2019

  12. [12]

    Matching anything by segmenting anything,

    S. Li, L. Ke, M. Danelljan, L. Piccinelli, M. Segu, L. Van Gool, and F. Yu, “Matching anything by segmenting anything,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  13. [13]

    SAM 2: Segment anything in images and videos,

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. R ¨adle, C. Rolland, L. Gustafsonet al., “SAM 2: Segment anything in images and videos,”arXiv preprint arXiv:2408.00714, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    Sam2long: Enhancing sam 2 for long video segmentation with a training-free memory tree,

    S. Ding, R. Qian, X. Dong, P. Zhang, Y . Zang, Y . Cao, Y . Guo, D. Lin, and J. Wang, “Sam2long: Enhancing sam 2 for long video segmentation with a training-free memory tree,” 2025. [Online]. Available: https://arxiv.org/abs/2410.16268

    Pith/arXiv arXiv 2025

  15. [15]

    A distractor-aware memory for visual object tracking with sam2,

    J. Videnovic, A. Lukezic, and M. Kristan, “A distractor-aware memory for visual object tracking with sam2,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 24 255–24 264

    2025

  16. [16]

    Sim ´eoni, H

    O. Sim ´eoni, H. V . V o, M. Seitzer, F. Baldassarre, M. Oquab, C. Jose, V . Khalidov, M. Szafraniec, S. Yi, M. Ramamonjisoa, F. Massa, D. Haz- iza, L. Wehrstedt, J. Wang, T. Darcet, T. Moutakanni, L. Sentana, C. Roberts, A. Vedaldi, J. Tolan, J. Brandt, C. Couprie, J. Mairal, H. J ´egou, P. Labatut, and P. Bojanowski, “DINOv3,”arXiv preprint arXiv:2508.1...

    Pith/arXiv arXiv 2025

  17. [17]

    ScanNet++: A high- fidelity dataset of 3d indoor scenes,

    C. Yeshwanth, Y .-C. Liu, M. Nießner, and A. Dai, “ScanNet++: A high- fidelity dataset of 3d indoor scenes,” inIEEE International Conference on Computer Vision (ICCV), 2023

    2023

  18. [18]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Loet al., “Segment anything,” inIEEE International Conference on Computer Vision (ICCV), 2023

    2023

  19. [19]

    TAO: A large-scale benchmark for tracking any object,

    A. Dave, T. Khurana, P. Tokmakov, C. Schmid, and D. Ramanan, “TAO: A large-scale benchmark for tracking any object,” inEuropean Conference on Computer Vision (ECCV), 2020

    2020

  20. [20]

    3am: Segment anything with geometric consistency in videos,

    Y .-C. Sun, C. Sun, C.-Y . Lin, F.-E. Yang, M.-H. Chen, Y .-Y . Lin, and Y .- L. Liu, “3am: Segment anything with geometric consistency in videos,” arXiv preprint arXiv:2601.08831, 2026

    Pith/arXiv arXiv 2026

  21. [21]

    Grounding image matching in 3d with MASt3R,

    V . Leroy, Y . Cabon, and J. Revaud, “Grounding image matching in 3d with MASt3R,” inEuropean Conference on Computer Vision (ECCV), 2024

    2024

  22. [22]

    Online segment any 3d thing as instance tracking,

    H. Wang, C. Zijian, J. Gao, Y . Zhang, W. Hu, K. Wang, and Z. Zhang, “Online segment any 3d thing as instance tracking,”Advances in Neural Information Processing Systems, vol. 38, pp. 49 076–49 095, 2026

    2026

  23. [23]

    Embodiedsam: Online segment any 3d thing in real time,

    X. Xu, H. Chen, L. Zhao, Z. Wang, J. Zhou, and J. Lu, “Embodiedsam: Online segment any 3d thing in real time,” inInternational Conference on Learning Representations, vol. 2025, 2025, pp. 38 020–38 035

    2025

  24. [24]

    Sam2object: Consolidating view consistency via sam2 for zero-shot 3d instance segmentation,

    J. Zhao, J. Zhuo, J. Chen, and H. Ma, “Sam2object: Consolidating view consistency via sam2 for zero-shot 3d instance segmentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2025, pp. 19 325–19 334

    2025

  25. [25]

    Any3dis: Class-agnostic 3d instance segmentation by 2d mask tracking,

    P. Nguyen, M. Luu, A. Tran, C. Pham, and K. Nguyen, “Any3dis: Class-agnostic 3d instance segmentation by 2d mask tracking,” 2024. [Online]. Available: https://arxiv.org/abs/2411.16183

    Pith/arXiv arXiv 2024

  26. [26]

    ScanNet: Richly-annotated 3d reconstructions of indoor scenes,

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, and M. Nießner, “ScanNet: Richly-annotated 3d reconstructions of indoor scenes,” inIEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2017

    2017

  27. [27]

    YOLOv11: An overview of the key architectural enhancements,

    R. Khanam and M. Hussain, “YOLOv11: An overview of the key architectural enhancements,”arXiv preprint arXiv:2410.17725, 2024. APPENDIXA THRESHOLDPARAMETERS REMIND is a heuristic-driven system whose behavior is governed by a large collection of scalar thresholds, each introduced at its first use in the main text. Table VIII con- solidates the most influent...

    Pith/arXiv arXiv 2024

This paper was first reviewed by grok-4.5 on July 13, 2026.