Pith. sign in

REVIEW 3 major objections 6 minor 66 references

PLOT: Pseudo-Labeling via Object Tracking for Monocular 3D Object Detection

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read PLOT generates 3D labels from monocular video alone—tracking points, recovering camera motion, and fusing per-object depth clouds—so a downstream detector reaches KITTI Car AP3D@0.3 of 80.48/60.83/51.49, matching fully supervised training.

desk verdict A genuinely new pose-free pseudo-labeling pipeline with unusually thorough benchmark evidence; the in-the-wild claims outrun the evidence and the scale-activation rule needs a quantitative definition, but the paper deserves serious review. read the letter →

arxiv 2507.02393 v2 pith:MSWDFBFT submitted 2025-07-03 cs.CV cs.GR

Seokyeong Lee , Sithu Aung , Junyong Choi , Seungryong Kim , Ig-Jae Kim , Junghyun Cho This is my paper
classification cs.CVcs.GR
keywords Pseudo-labelingMonocular3DobjectdetectiontrackingDensepointPseudo-LiDARfusionGlobalmemoryCameramotionestimationVideo-basedperception
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

PLOT claims that 3D object annotations can be produced directly from monocular video alone—no LiDAR, stereo, IMU, known camera poses, or model retraining. The framework tracks object and background points across frames, recovers relative camera motion by aligning depth-lifted background trajectories, and fuses each object's per-frame pseudo-LiDAR observations into a completed shape, with a global object memory holding identities stable through occlusion and detector noise. The authors verify the claim by training a standard monocular detector on the generated labels and reporting AP3D@0.3 of 80.48/60.83/51.49 on the KITTI Car validation set, comparable to the same detector trained on ground-truth 3D labels and above every prior pseudo-labeling method. This matters because 3D label scarcity is what confines monocular 3D detection to curated, sensor-equipped benchmarks; if the claim holds, abundant ordinary video becomes a source of 3D supervision for driving, surveillance, handheld, and in-the-wild scenes.

What carries the argument

Trajectory-guided shape fusion is the load-bearing mechanism. Monocular depth maps lift tracked 2D masks into per-frame pseudo-LiDAR point clouds; background point trajectories are aligned by Procrustes registration (Eq. 4) to recover relative camera rotation and translation, with an optional scale factor that compensates temporal depth drift, while object point trajectories are registered the same way (Eq. 5) to recover per-object motion. The recovered relative transforms warp every observation of an object into a single reference frame, fusing partial views into one completed point cloud (Eq. 9); the denser geometry stabilizes PCA-based orientation, size, and center estimation compared with single-frame or sparsely aggregated observations. The Global Object Memory is a second mechanism: it maintains persistent per-object entries, discards unstable instances, refines attributes with dominant evidence, and merges redundant tracks, directly addressing the two failure modes—detection dropouts and identity switches—that would otherwise fragment the fused shape.

What would settle it

A direct check: on any KITTI sequence, compute the relative camera poses from Eq. 4 and compare them against the benchmark's ground-truth trajectory derived from its calibration and GPS data; if the estimated rotations and translations deviate beyond a few degrees and decimeters, the fused object point clouds inherit the error, and substituting the ground-truth relative poses should measurably change downstream box AP. The paper's own stability table provides the sharper variant: injecting non-uniform temporal scale drift into the depth maps should push box AP at IoU 0.3 back toward the single-frame baseline, which is the signature of the assumption failing.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that temporal correspondence—not extra sensors—resolves the depth–scale ambiguity of monocular 3D detection. A dense point tracker propagates per-frame detection masks across time, giving object points and background points stable identities; lifting those trajectories to 3D with a monocular metric depth estimator yields frame-wise pseudo-LiDAR point clouds. Background trajectories are registered by Procrustes alignment (Eq. 4) to estimate the relative camera pose between each frame and a reference frame, and object trajectories receive the same registration (Eq. 5) to separate object motion from ego-motion. These relative transforms warp every observation of an object into one coordinate frame, producing a completed pseudo-LiDAR (Eq. 9) that is reprojected back to each frame so center, size, and orientation are annotated consistently; the Global Object Memory merges, refines, or drops object hypotheses so that dropped detections and identity switches do not corrupt the fusion. The central validation is a downstream detector trained on the labels reaching AP3D@0.3 of 80.48/60.83/51.49 on KITTI Car (against 79.72/65.87/58.83 for the same detector with full supervision), and the raw labels—used with no training at all—already surpassing single-image pseudo-labeling and several weakly supervised baselines.

Load-bearing premise

The load-bearing premise is that the relative camera poses estimated by Procrustes alignment of depth-lifted background point trajectories (Eq. 4) are metrically accurate and scale-consistent across frames; if the monocular depth is biased or drifts temporally, the fused pseudo-LiDAR inherits the error and the object center, size, and orientation estimates degrade accordingly.

Editorial extensions

If this is right

  • Aggregating more frames improves label quality monotonically: fusing a 20-frame window raises average AP by about 1.7x over single-frame pseudo-labels, at roughly 2.8 seconds per frame for 20 objects.
  • The raw pseudo-labels are useful with zero training: run directly as a labeler, they already outperform the single-image pseudo-labeler and weakly supervised baselines on KITTI before any downstream detector is trained.
  • The pipeline transfers beyond driving: qualitative results on surveillance, handheld, crowd-sourced, and phone-captured videos show stable boxes where open-set detectors and prior pseudo-labelers fail or degrade.
  • On Waymo-Open, labels generated under night and rain improve downstream AP over fully supervised MonoDETR in ranges beyond 30 m and recover valid instances missing from the ground-truth annotation.
  • Pedestrian labels stay accurate despite non-rigid motion and occlusion, indicating that the fusion approach tolerates violations of the rigid-shape assumption used for cars.

Reading between the lines

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

  • Every generated label inherits the metric scale and bias of the chosen monocular depth estimator, so the ceiling of the whole pipeline is the depth model; pairing PLOT with a depth estimator whose per-range bias is characterized would let users attach error bars to each label.
  • The same tracking-plus-fusion recipe should apply to class-agnostic instance segmentation: any category a 2D open-vocabulary detector can segment becomes a candidate for 3D pseudo-labeling, which is exactly where single-image LLM-prior methods break down.
  • The window-based fusion and the qualitative criterion for activating the Eq. (4) scale factor are underspecified; an automatic scale-drift test based on the depth estimator's own statistics, plus overlapping windows to redeem boundary association failures, would make the method fully reproducible and reduce the edge-case association errors the authors report.
  • Because the method only needs relative poses, it could label sequences captured by cameras with unknown and even changing intrinsics, as long as the depth model supplies calibration, suggesting a path to bootstrapping 3D benchmarks for egocentric and body-mounted video.
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

3 major / 6 minor

Summary. PLOT proposes a training-free pipeline that generates 3D pseudo-labels for monocular object detection from monocular video. The method uses off-the-shelf 2D detection/segmentation (Grounded-SAM), dense point tracking (AllTracker), and metric depth estimation (UniDepth or MoGe2). Object and background point trajectories are lifted to 3D and aligned by Procrustes registration to recover relative camera/object motion without known poses. A Global Object Memory (GOM) maintains object identities across frames, and trajectory-guided fusion aggregates per-frame pseudo-LiDAR into completed object shapes. A downstream MonoDETR is trained on the resulting labels and evaluated on KITTI, KITTI-360, and Waymo, with qualitative results on in-the-wild videos. The paper reports AP3D@0.3 of 80.48/60.83/51.49 on KITTI Car validation, comparable to fully supervised baselines and ahead of prior pseudo-labeling methods.

Significance. If the reported results hold, this is a valuable contribution: it offers a plausible route to 3D supervision from monocular video without auxiliary sensors or explicit camera poses, which would broaden the applicability of monocular 3D detection to unconstrained settings. The experimental apparatus is a genuine strength: the paper includes ablations over temporal window size, GOM, camera motion modeling, estimated vs. ground-truth intrinsics, and a carefully designed noise-injection study in the supplementary material. The raw pseudo-label analysis (Table 7) and standalone-detector experiment (Table 6) provide transparency beyond the final downstream AP numbers. The core geometric mechanism—Procrustes alignment of tracked points followed by multi-frame shape fusion—is coherent and standard enough to be credible, and the authors do not fit the pseudo-labels to the 3D ground truth used for evaluation. The main weaknesses are an underspecified scale-activation rule that is load-bearing for metric accuracy, and a mismatch between the paper's strong unconstrained-domain claims and the purely qualitative evidence offered for in-the-wild videos.

major comments (3)
  1. [Sec. 3.2, Eq. (4); Sec. 3.4, Eq. (9); Supp. A1.2] The optional scale factors s_t^c and s_t are load-bearing for the metric correctness of the pseudo-labels. The paper states that s_t^c is activated only when monocular depth exhibits 'noticeable temporal scale drift or flicker,' but no quantitative activation criterion is given, and the supplementary hyperparameter list (A1.2) does not specify when the similarity transform is used instead of the rigid one. Because Eq. (9) applies the same scale to the fused object points, object dimensions and centers inherit whatever scale the Procrustes fit chooses. The paper's own robustness study (Table A1) shows that temporal scale noise degrades Easy AP3D@0.3 from 20.89 to 14.14 under SE(3) alignment and that Sim(3) only partially recovers the loss. The main benchmark numbers use ground-truth intrinsics and UniDepth, where the default s=1 may be safe, but the in-the-wild results use estimated intrinsics and a different depth model; the method as currently specified is not reproducible without the activation rule. Please define the exact rule and report which configuration (rigid or similarity) is used for every experimental result.
  2. [Sec. 4.2; Supp. A5.5; Supp. Table A3] The in-the-wild evaluation is entirely qualitative. Section 4.2 and Figure 7 show example predictions, and Figure A10 shows BEV overlays, but there is no quantitative 3D metric, no ground-truth comparison, and no recall/precision measurement for MOT17, MOSE, DIVOTrack, GMOT-40, or Pexels. This matters because the abstract claims 'strong generalization' and 'proving its effectiveness across diverse and unconstrained domains,' which is exactly the regime where the no-auxiliary-sensor claim is most consequential and where the pipeline uses estimated intrinsics and MoGe2 depth instead of the ground-truth intrinsics and UniDepth used in the driving benchmarks. The authors' own Table A3 shows that replacing ground-truth intrinsics with UniDepth-estimated intrinsics on KITTI drops Easy AP3D@0.3 from 80.48 to 70.44, indicating substantial sensitivity to the components that differ in the in-the-wild setting. I ask either for quantitative evaluation on non-driving 3D benchmarks or for a carefully scoped claim that is limited to the driving benchmarks that are actually measured.
  3. [Sec. 4.3, Table 7; Supp. A3] Table 7 reports 'pseudo-label accuracy' using 2D AP, AOE, and ASE, but not a 3D localization metric such as center distance error, so the claim that the raw pseudo-labels are 'reliable' in 3D is only partially supported by this table. The 2D AP is largely inherited from the GSAM detector, and the AOE/ASE numbers are computed after some object-to-ground-truth association procedure that is not described in the paper. Since the downstream MonoDETR results (Tables 1-4) are the strongest evidence for label quality, the paper should either clarify the association protocol for Table 7 or add a direct 3D center/size error evaluation against KITTI ground truth.
minor comments (6)
  1. [Table 6] The header 'Train GOM' is ambiguous: the first KITTI row appears to be raw pseudo-labels with GOM but without MonoDETR training, while the second row includes training. Please label the rows explicitly (e.g., 'MonoDETR training: yes/no').
  2. [Supp. Table A1] The row labels 'Range Patch Temporal TinyDepth Pose' are difficult to parse; please expand the column names and explain the exact noise models used for each perturbation.
  3. [Eq. (7)] The notation p^i(z) and p^i(x) is used for displacement, but the coordinate frame of the world-aligned trajectory from Eq. (6) is not defined; please state explicitly which axis is the ground-plane forward axis.
  4. [Sec. 3.1] The paper alternates between 'metric depth maps' and 'monocular depth predictions'; since the optional scale factor in Eq. (4) interacts with the definition of metric depth, please use consistent terminology and clarify whether the depth is intended to be metric before the optional scale is applied.
  5. [Supp. A1.1] The claim of 'partial source code for reproducibility' should specify which components are included (e.g., GOM, Procrustes alignment, or the full label generation loop); without this, reproducing Table 1 would require re-implementing the full pipeline.
  6. [Supp. A2.2] There is a typo, 'simliar', in the description of using adjacent frames for KITTI; please correct it.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PLOT's pseudo-labels come from an independent geometric fusion pipeline and are evaluated against external 3D ground truth, not from fitted labels or self-citation.

full rationale

PLOT's derivation chain is self-contained. Frame-wise depth maps and 2D point tracks are inputs; Eq. (4) and Eq. (5) solve Procrustes registrations on background and object correspondences, and Eq. (9) aggregates the transformed pseudo-LiDAR points. Object attributes (center, size, orientation) are then read off the fused point cloud, while Tab. 1 and Tab. 7 evaluate these labels against KITTI/Waymo 3D ground truth that is not used in label generation. Hyperparameters (2 m motion threshold, 20-frame window, visibility threshold) are set from general video properties rather than fit to the validation labels. The optional scale term s_t^c in Eq. (4) is activated by an underspecified qualitative criterion, and depth scale errors could propagate into object dimensions; this is a correctness and reproducibility limitation, not a circular reduction, because the output is not an input or a fitted quantity renamed as a prediction. No load-bearing self-citations appear: the only likely in-house citation (Seurat [8]) is used in related-work contrast, not as evidence for any derivation. The paper is therefore self-contained against external benchmarks, so no circularity is found.

Assumptions & free parameters 6 free parameters · 6 assumptions · 1 invented entities

The pipeline does not introduce new physical entities; its load-bearing inputs are pre-trained neural models and hand-set thresholds. The main hand-set parameters are the optional scale factors in motion estimation, the motion and association thresholds, and the tracking window length. The axioms are standard rigid and static-scene assumptions plus reliance on metric depth and consistent point tracks.

free parameters (6)
  • Optional scale factors s_t^c and s_t (Eq. 4, 5) = 1 by default; optimized when depth shows scale drift
    Activated only when monocular depth is judged to have temporal scale drift or flicker; the activation rule is qualitative, so this parameter is underspecified.
  • Motion classification threshold = 2 m displacement between adjacent frames
    Objects with displacement above 2 m are treated as dynamic (Supp A1.2); hand-set and affects orientation estimation strategy.
  • Tracking window length = 20 frames
    Adopted in the main experiments based on the accuracy-versus-time tradeoff in Tab. 5; hand-selected.
  • Hungarian matching IoU threshold = 0.4
    Threshold for associating tracked masks to predicted masks (Eq. 2, Supp A1.2); hand-set.
  • Point-track visibility threshold = 0.6
    Used to suppress unreliable correspondences in camera motion estimation (Supp A1.2); hand-set.
  • Background depth filter = 50 m cutoff
    Background points beyond 50 m are excluded from camera motion estimation to suppress noisy depth (Sec. 3.2, Supp A1.2); follows common practice but is a hand-set cutoff.
assumptions (6)
  • domain assumption Monocular depth estimators (UniDepth, MoGe2) provide reliable metric scale per frame.
    The entire pipeline lifts 2D points to 3D using these depth maps (Sec. 3.1, 3.2); if depth scale is wrong, the fused pseudo-LiDAR is wrong.
  • domain assumption Background is static, so background point trajectories constrain ego-motion.
    Camera motion is estimated by Procrustes alignment on background tracks (Eq. 4); moving background objects would corrupt the fit.
  • domain assumption Objects are rigid during the fusion window.
    A single rigid transformation per object per frame is used to align pseudo-LiDAR (Eq. 5, Eq. 9); non-rigid deformation (e.g., pedestrians) is handled implicitly but can bias shape.
  • ad hoc to paper The principal axis (highest variance) of the fused object point cloud is the object's longitudinal axis.
    Static-object orientation is computed via PCA on the completed pseudo-LiDAR (Sec. 3.2); this heuristic fails for symmetric or non-canonical shapes.
  • domain assumption Camera intrinsics are known or accurately estimated by the depth model.
    Depth maps are unprojected to 3D using intrinsics (Supp A1.1); Tab. A3 shows only mild degradation with estimated intrinsics, so this is a mild assumption.
  • domain assumption Point tracker provides dense, consistent correspondences across frames.
    All motion estimation and association rest on AllTracker's tracks (Sec. 3.1); if tracks are spurious, the associations and alignments break.
invented entities (1)
  • Global Object Memory (GOM)
    purpose: Maintains persistent object identities, discards unstable instances, and merges redundant entries to reduce ID switches and recover missed detections.
    GOM is an algorithmic construct proposed in Sec. 3.3; its benefit is demonstrated by ablations (Tab. 6) within the paper, but it has no external falsifiable handle beyond those experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PLOT: Pseudo-Labeling via Object Tracking for Monocular 3D Object Detection." pith.science (2026). https://pith.science/paper/MSWDFBFT

@misc{pith2026250702393,
  author       = {Pith},
  title        = {Pith review of: PLOT: Pseudo-Labeling via Object Tracking for Monocular 3D Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MSWDFBFT}},
  note         = {Machine review of arXiv:2507.02393}
}
read the original abstract

Monocular 3D object detection is crucial for scalable perception across fields like autonomous driving, robotics, and surveillance. However, progress is hindered by limited 3D annotations and the inherent ambiguity of single-image geometry. Existing methods often rely on strong geometric assumptions or carefully curated datasets, which limit their applicability to real-world scenarios. In this paper, we present PLOT (Pseudo-Labeling via Object Tracking), a framework that generates 3D annotations from monocular videos without auxiliary sensors or model retraining. PLOT tracks object and background trajectories to estimate camera motion and perform object association in pose-unknown settings. These trajectories provide point correspondences that align frame-wise pseudo-LiDARs, which are then fused via simple optimization into a unified object shape robust to occlusion and viewpoint shifts. Recognizing temporal coherence as a fundamental requirement for reliable shape fusion and video perception, we design a global object memory that preserves consistent object identities across frames. PLOT achieves robust annotation quality and strong generalization on both M3OD video benchmarks and in-the-wild videos, proving its effectiveness across diverse and unconstrained domains. Project page: https://plot-eccv.github.io.

Figures

Figures reproduced from arXiv: 2507.02393 by the authors.

Figure 1
Figure 1. Zero-shot estimation on MOT17 [36] and Pexels [9]: Predictions from (top) an open-set M3OD model [59] and (bottom) our pseudo-labeling method. Our method performs well on out-of-domain camera views (zoom in for a better view). lack of direct geometric supervision, which has led most methods to rely on curated benchmarks collected in sensor-rich environments, such as autonomous driving [7, 12, 51] and indoor scenes [… view at source ↗
Figure 2
Figure 2. PLOT generates accurate 3D labels directly from monocular videos without requiring auxiliary sensors or training, as illustrated in (a) qualitative results across diverse scenarios. Furthermore, (b) our object tracking and aggregation pipeline pro￾duces shape-complete pseudo-LiDARs, yielding BEV maps comparable to ground truth and can identify miss-labeled objects (marked with a red star). cent work [50] attempts to… view at source ↗
Figure 3
Figure 3. Overall architecture of PLOT. Given monocular videos, we extract 2D detec￾tions and depth, and track points to obtain temporally grounded correspondences. These are used to estimate relative poses and camera motions across frames, enabling shape fusion and orientation estimation. A global object association module refines trajectories and recovers missing instances. Finally, completed pseudo-LiDARs are re￾projected … view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Global object association pipeline. Our memory-based object association con￾solidates noisy frame-level predictions into globally consistent tracks. Although 2D point trackers [15, 24] and detectors [47] are effective per frame, occlusions and clutter still cause misse…
Figure 6
Figure 6. Figure 6: Qualitative comparison on Waymo-Open. OVM3D-Det [21] 3D-MOOD [59] PLOT (Ours) [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Qualitative comparisons in the wild. real-world complexity. PLOT achieves the best performance in distant regions (>30m), surpassing fully supervised baselines, and shows superior APBEV across nearly all ranges. As shown in [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: BEV comparison between the detection results of raw pseudo-labels and the detector trained with it, on KITTI [12]. estimates with most falls back to priors of LLM-driven fixed objects, due to single image observations. Such in-the-wild evaluations are particularly impo…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

66 extracted references · 45 canonical work pages

  1. [1]

    arXiv preprint arXiv:2303.08774 (2023)

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    arXiv 2023

  2. [2]

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

    Bai, H., Cheng, W., Chu, P., Liu, J., Zhang, K., Ling, H.: Gmot-40: A benchmark for generic multiple object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6719–6728 (2021)

    work page 2021

  3. [3]

    arXiv preprint arXiv:2111.08897 (2021)

    Baruch, G., Chen, Z., Dehghan, A., Dimry, T., Feigin, Y., Fu, P., Gebauer, T., Joffe, B., Kurz, D., Schwartz, A., et al.: Arkitscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. arXiv preprint arXiv:2111.08897 (2021)

    arXiv 2021

  4. [4]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Brazil, G., Kumar, A., Straub, J., Ravi, N., Johnson, J., Gkioxari, G.: Omni3d: A large benchmark and model for 3d object detection in the wild. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13154– 13164 (2023)

    work page 2023

  5. [5]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Brazil, G., Liu, X.: M3d-rpn: Monocular 3d region proposal network for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9287–9296 (2019)

    work page 2019

  6. [6]

    In: Computer Vision–ECCV 2020: 16th European Confer- ence, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXIII 16

    Brazil, G., Pons-Moll, G., Liu, X., Schiele, B.: Kinematic 3d object detection in monocular video. In: Computer Vision–ECCV 2020: 16th European Confer- ence, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXIII 16. pp. 135–152. Springer (2020)

    work page 2020

  7. [7]

    In: CVPR (2020)

    Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020)

    2020

  8. [8]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2025)

    Cho, S., Huang, J., Kim, S., Lee, J.Y.: Seurat: From moving points to depth. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2025)

    work page 2025

Show all 66 references
  1. [9]

    pexels.com/, accessed: 2025-03-04

    Contributors, P.: Pexels - free stock photos and videos (2025),https://www. pexels.com/, accessed: 2025-03-04

    2025

  2. [10]

    In: Proc

    Dai,A.,Chang,A.X.,Savva,M.,Halber,M.,Funkhouser,T.,Nießner,M.:Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proc. Computer Vision and Pattern Recognition (CVPR), IEEE (2017)

    2017

  3. [11]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Ding, H., Liu, C., He, S., Jiang, X., Torr, P.H., Bai, S.: Mose: A new dataset for video object segmentation in complex scenes. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 20224–20234 (2023)

    2023

  4. [12]

    In: 2012 IEEE conference on computer vision and pattern recognition

    Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3354–3361. IEEE (2012)

    2012

  5. [13]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    Han, W., Tao, R., Ling, H., Shen, J.: Weakly supervised monocular 3d object detec- tion by spatial-temporal view consistency. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    2024

  6. [14]

    International Journal of Computer Vision132(4), 1075–1090 (2024)

    Hao, S., Liu, P., Zhan, Y., Jin, K., Liu, Z., Song, M., Hwang, J.N., Wang, G.: Divo- track: A novel dataset and baseline method for cross-view multi-object tracking in diverse open scenes. International Journal of Computer Vision132(4), 1075–1090 (2024)

    2024

  7. [15]

    In: ICCV (2025) PLOT 29

    Harley, A.W., You, Y., Sun, X., Zheng, Y., Raghuraman, N., Gu, Y., Liang, S., Chu, W.H., Dave, A., Tokmakov, P., You, S., Ambrus, R., Fragkiadaki, K., Guibas, L.J.: AllTracker: Efficient dense point tracking at high resolution. In: ICCV (2025) PLOT 29

    2025

  8. [16]

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

    Hu,J.S.K.,Kuai,T.,Waslander,S.L.:Pointdensity-awarevoxelsforlidar3dobject detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8469–8478 (June 2022)

    2022

  9. [17]

    In: arXiv preprint (2024)

    Hu, M., Yin, W., Zhang, C., Cai, Z., Long, X., Chen, H., Wang, K., Yu, G., Shen, C., Shen, S.: Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. In: arXiv preprint (2024)

    2024

  10. [18]

    In: NVIDIA Research Whitepapers (2025)

    Huang, J., Zhou, Q., Rabeti, H., Korovko, A., Ling, H., Ren, X., Shen, T., Gao, J., Slepichev, D., Lin, C.H., Ren, J., Xie, K., Biswas, J., Leal-Taixe, L., Fidler, S.: Vipe: Video pose engine for 3d geometric perception. In: NVIDIA Research Whitepapers (2025)

    2025

  11. [19]

    In: ECCV (2024)

    Huang, K.C., Tsai, Y.H., Yang, M.H.: Weakly supervised 3d object detection via multi-level visual guidance. In: ECCV (2024)

    2024

  12. [20]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Huang,K.C.,Wu,T.H.,Su,H.T.,Hsu,W.H.:Monodtr:Monocular3dobjectdetec- tion with depth-aware transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4012–4021 (2022)

    2022

  13. [21]

    In: The Thirty-eighth Annual Conference on Neural Information Processing Systems (2024)

    Huang, R., Zheng, H., Wang, Y., Xia, Z., Pavone, M., Huang, G.: Training an open- vocabulary monocular 3d detection model without 3d data. In: The Thirty-eighth Annual Conference on Neural Information Processing Systems (2024)

    2024

  14. [22]

    In: Thirty-seventh Conference on Neural Information Processing Systems (2023)

    Jia, J., Li, Z., Shi, Y.: Monouni: A unified vehicle and infrastructure-side monoc- ular 3d object detection network with sufficient depth clues. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)

    2023

  15. [23]

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

    Jiang, X., Jin, S., Lu, L., Zhang, X., Lu, S.: Weakly supervised monocular 3d detection with a single-view image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10508–10518 (2024)

    2024

  16. [24]

    In: European Conference on Computer Vision

    Karaev, N., Rocco, I., Graham, B., Neverova, N., Vedaldi, A., Rupprecht, C.: Cotracker: It is better to track together. In: European Conference on Computer Vision. pp. 18–35. Springer (2025)

    2025

  17. [25]

    Naval research logistics quarterly2(1-2), 83–97 (1955)

    Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly2(1-2), 83–97 (1955)

    1955

  18. [26]

    In: European Conference on Computer Vision

    Kumar, A., Brazil, G., Corona, E., Parchami, A., Liu, X.: Deviant: Depth equiv- ariant network for monocular 3d object detection. In: European Conference on Computer Vision. pp. 664–683. Springer (2022)

    2022

  19. [27]

    In: European Conference on Computer Vision

    Leroy, V., Cabon, Y., Revaud, J.: Grounding image matching in 3d with mast3r. In: European Conference on Computer Vision. pp. 71–91. Springer (2024)

    2024

  20. [28]

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

    Li, Y., Wu, C.Y., Fan, H., Mangalam, K., Xiong, B., Malik, J., Feichtenhofer, C.: Mvitv2: Improved multiscale vision transformers for classification and detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4804–4814 (June 2022)

    2022

  21. [29]

    Liao, Y., Xie, J., Geiger, A.: KITTI-360: A novel dataset and benchmarks for urbansceneunderstandingin2dand3d.PatternAnalysisandMachineIntelligence (PAMI) (2022)

    2022

  22. [30]

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

    Liu, X., Zheng, C., Cheng, K.B., Xue, N., Qi, G.J., Wu, T.: Monocular 3d object detection with bounding box denoising in 3d by perceiver. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6436–6446 (2023)

    2023

  23. [31]

    Liu, Z., Sakuma, H., Okutomi, M.: Vsrd: Instance-aware volumetric silhouette ren- deringforweaklysupervised3dobjectdetection.In:ProceedingsoftheIEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17354–17363 (2024)

    2024

  24. [32]

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

    Liu, Z., Zhou, D., Lu, F., Fang, J., Zhang, L.: Autoshape: Real-time shape-aware monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15641–15650 (2021) 30 S. Lee et al

    2021

  25. [33]

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

    Lu, Y., Ma, X., Yang, L., Zhang, T., Liu, Y., Chu, Q., Yan, J., Ouyang, W.: Geometry uncertainty projection network for monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3111–3121 (2021)

    2021

  26. [34]

    In: Inter- national Conference on Computer Analysis of Images and Patterns

    Luo, B., Hancock, E.R.: Procrustes alignment with the em algorithm. In: Inter- national Conference on Computer Analysis of Images and Patterns. pp. 623–631. Springer (1999)

    1999

  27. [35]

    IEEE transactions on pattern analysis and machine intelligence45(1), 1320–1327 (2022)

    Maninis, K.K., Popov, S., Nießner, M., Ferrari, V.: Vid2cad: Cad model alignment using multi-view constraints from videos. IEEE transactions on pattern analysis and machine intelligence45(1), 1320–1327 (2022)

    2022

  28. [36]

    arXiv:1603.00831 [cs] (Mar 2016),http://arxiv.org/ abs/1603.00831, arXiv: 1603.00831

    Milan, A., Leal-Taixé, L., Reid, I., Roth, S., Schindler, K.: MOT16: A benchmark for multi-object tracking. arXiv:1603.00831 [cs] (Mar 2016),http://arxiv.org/ abs/1603.00831, arXiv: 1603.00831

    2016 arXiv

  29. [37]

    arXiv preprint arXiv:2304.07193 (2023)

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)

    2023 arXiv

  30. [38]

    In: The Twelfth International Conference on Learning Representations (2024)

    Pan, T.Y., Ma, C., Chen, T., Phoo, C.P., Luo, K.Z., You, Y., Campbell, M., Weinberger, K.Q., Hariharan, B., Chao, W.L.: Pre-training liDAR-based 3d ob- ject detectors through colorization. In: The Twelfth International Conference on Learning Representations (2024)

    2024

  31. [39]

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

    Parihar, R., Sarkar, S., Vora, S., Kundu, J.N., Babu, R.V.: Monoplace3d: Learning 3d-aware object placement for 3d monocular detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6531–6541 (June 2025)

    2025

  32. [40]

    In: European Conference on Computer Vision

    Peng, L., Wu, X., Yang, Z., Liu, H., Cai, D.: Did-m3d: Decoupling instance depth for monocular 3d object detection. In: European Conference on Computer Vision. pp. 71–88. Springer (2022)

    2022

  33. [41]

    arXiv preprint arXiv:2203.08332 (2022)

    Peng, L., Yan, S., Wu, B., Yang, Z., He, X., Cai, D.: Weakm3d: Towards weakly supervised monocular 3d object detection. arXiv preprint arXiv:2203.08332 (2022)

    2022 arXiv

  34. [42]

    In: CVPR

    Piccinelli, L., Yang, Y.H., Sakaridis, C., Segu, M., Li, S., Van Gool, L., Yu, F.: Unidepth: Universal monocular metric depth estimation. In: CVPR. pp. 10106– 10116 (2024)

    2024

  35. [43]

    In: Proceedings of the 28th ACM International Conference on Multimedia

    Qin, Z., Wang, J., Lu, Y.: Weakly supervised 3d object detection from point clouds. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 4144–4152 (2020)

    2020

  36. [44]

    IEEE transactions on pattern analysis and machine intelligence 44(9), 5170–5184 (2021)

    Qin, Z., Wang, J., Lu, Y.: Monogrnet: A general framework for monocular 3d object detection. IEEE transactions on pattern analysis and machine intelligence 44(9), 5170–5184 (2021)

    2021

  37. [45]

    arXiv preprint arXiv:2408.00714 (2024)

    Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., et al.: Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714 (2024)

    2024 arXiv

  38. [46]

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

    Reading, C., Harakeh, A., Chae, J., Waslander, S.L.: Categorical depth distribution network for monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8555–8564 (2021)

    2021

  39. [47]

    arXiv preprint arXiv:2401.14159 (2024)

    Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., et al.: Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159 (2024)

    2024 arXiv

  40. [48]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Shi, S., Wang, X., Li, H.: Pointrcnn: 3d object proposal generation and detection from point cloud. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 770–779 (2019) PLOT 31

    2019

  41. [49]

    In: Proceedings of the IEEE/CVF in- ternational conference on computer vision

    Simonelli, A., Bulo, S.R., Porzi, L., López-Antequera, M., Kontschieder, P.: Dis- entangling monocular 3d object detection. In: Proceedings of the IEEE/CVF in- ternational conference on computer vision. pp. 1991–1999 (2019)

    2019

  42. [50]

    arXiv preprint arXiv:2501.09481 (2025)

    Skvrna, J., Neumann, L.: Monosowa: Scalable monocular 3d object detector with- out human annotations. arXiv preprint arXiv:2501.09481 (2025)

    2025 arXiv

  43. [51]

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

    Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalab...

    2020

  44. [52]

    Tao, R., Han, W., Qiu, Z., Xu, C.Z., Shen, J.: Weakly supervised monocular 3d objectdetectionusingmulti-viewprojectionanddirectionconsistency.In:Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17482–17492 (2023)

    2023

  45. [53]

    Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test- timeadaptationbyentropyminimization.In:InternationalConferenceonLearning Representations (2021)

    2021

  46. [54]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Wang, J., Chen, M., Karaev, N., Vedaldi, A., Rupprecht, C., Novotny, D.: Vggt: Visual geometry grounded transformer. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 5294–5306 (2025)

    2025

  47. [55]

    In: International Conference on Computer Vision (ICCV) (2025)

    Wang, Q., Ye, V., Gao, H., Zeng, W., Austin, J., Li, Z., Kanazawa, A.: Shape of motion: 4d reconstruction from a single video. In: International Conference on Computer Vision (ICCV) (2025)

    2025

  48. [56]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Wang, Q., Zhang, Y., Holynski, A., Efros, A.A., Kanazawa, A.: Continuous 3d perception model with persistent state. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 10510–10522 (2025)

    2025

  49. [57]

    arXiv preprint arXiv:2507.02546 (2025)

    Wang, R., Xu, S., Dong, Y., Deng, Y., Xiang, J., Lv, Z., Sun, G., Tong, X., Yang, J.: Moge-2: Accurate monocular geometry with metric scale and sharp details. arXiv preprint arXiv:2507.02546 (2025)

    2025 arXiv

  50. [58]

    In: European Conference on Computer Vision

    Wang, T., Pang, J., Lin, D.: Monocular 3d object detection with depth from mo- tion. In: European Conference on Computer Vision. pp. 386–403. Springer (2022)

    2022

  51. [59]

    arXiv preprint arXiv:2507.23567 (2025)

    Yang, Y.H., Piccinelli, L., Segu, M., Li, S., Huang, R., Fu, Y., Pollefeys, M., Blum, H., Bauer, Z.: 3d-mood: Lifting 2d to 3d for monocular open-set object detection. arXiv preprint arXiv:2507.23567 (2025)

    2025 arXiv

  52. [60]

    arXiv preprint arXiv:2411.16833 (2024)

    Yao, J., Gu, H., Chen, X., Wang, J., Cheng, Z.: Open vocabulary monocular 3d object detection. arXiv preprint arXiv:2411.16833 (2024)

    2024

  53. [61]

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

    Zakharov, S., Kehl, W., Bhargava, A., Gaidon, A.: Autolabeling 3d objects with differentiable rendering of sdf shape priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12224–12233 (2020)

    2020

  54. [62]

    In: European Conference on Computer Vision

    Zhang, G., Fan, J., Chen, L., Zhang, Z., Lei, Z., Zhang, L.: General geometry-aware weakly supervised 3d object detection. In: European Conference on Computer Vision. pp. 290–309. Springer (2024)

    2024

  55. [63]

    arXiv preprint arXiv:2504.07958 (2025)

    Zhang, H., Jiang, H., Yao, Q., Sun, Y., Zhang, R., Zhao, H., Li, H., Zhu, H., Yang, Z.: Detect anything 3d in the wild. arXiv preprint arXiv:2504.07958 (2025)

    2025

  56. [64]

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

    Zhang, J., Li, J., Lin, X., Zhang, W., Tan, X., Han, J., Ding, E., Wang, J., Li, G.: Decoupled pseudo-labeling for semi-supervised monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16923–16932 (June...

    2024

  57. [65]

    In: International Conference on Learning Representations (2025)

    Zhang, J., Herrmann, C., Hur, J., Jampani, V., Darrell, T., Cole, F., Sun, D., Yang, M.H.: Monst3r: A simple approach for estimating geometry in the presence of motion. In: International Conference on Learning Representations (2025)

    2025

  58. [66]

    Zhang, R., Qiu, H., Wang, T., Guo, Z., Cui, Z., Qiao, Y., Li, H., Gao, P.: Monodetr: Depth-guidedtransformerformonocular3dobjectdetection.In:Proceedingsofthe IEEE/CVF International Conference on Computer Vision. pp. 9155–9166 (2023)

    2023

Pith tools

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