Pith. sign in

REVIEW 4 major objections 5 minor 53 references

Depth as Points: Center Point-based Depth Estimation

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read Monocular depth for driving can be estimated from an object's center point using a local conditional random field, the paper claims.

desk verdict A useful synthetic-data engineering idea is buried under an underspecified core module and inconsistent numbers; the central claim is not verifiable as written. read the letter →

arxiv 2504.18773 v1 pith:U7LNNRTW submitted 2025-04-26 cs.CV

classification cs.CV
keywords monoculardepthestimationcenterpointdetectionconditionalrandomfieldsautonomousdrivingvirtualdatasetCARLAkeypoint-basedlong-rangeperception
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

The paper argues that monocular depth estimation for autonomous driving should be object-centric rather than dense: locate each vehicle or pedestrian by its center point, then regress that center's depth directly. To make this work at long range, it introduces Center FC-CRFs, a fully-connected conditional random field restricted to the detection box, with the center acting as the anchor for feature and depth propagation. The paper also contributes a CARLA-based data generator and the VirDepth synthetic dataset, built in hours with KITTI-matching camera parameters. On VirDepth, Virtual KITTI 2, and KITTI subsets, it reports that CenterDepth reaches $\delta_1$ 0.989 and MRE 0.041 with the Hourglass-104 backbone and MAE 3.351 m in the 150-200 m band, at real-time inference speeds. If these results hold, the implication is that center-point depth can stay accurate where full-image depth estimators and monocular 3D detectors degrade on small distant targets.

What carries the argument

The load-bearing mechanism is the Center FC-CRFs module: a fully-connected conditional random field restricted to the object's bounding box, with the detected center point as the anchor node. Pairwise potentials tie each pixel's depth to the center depth with weights based on feature similarity (Eq. 4), and the depth map is obtained by minimizing the energy in Eq. 5. This module is what converts the center-point heatmap into a depth value per target, and the paper attributes the large gains of CenterDepth over its no-CRF baseline to it. The second mechanism is the virtual data generator: CARLA meshes are projected through KITTI-matched intrinsics and extrinsics to produce 2D boxes, 3D boxes, and dense depth labels automatically, creating VirDepth with more than 180,000 targets between 0 and 200 m.

What would settle it

Implement Center FC-CRFs exactly from the energy function in Eqs. (2)-(5) and train the same backbone on VirDepth with the module removed versus included; the paper gives no loss term or inference update rule for the energy minimization, so a reader cannot reproduce it. If a direct implementation does not reproduce the large gap between CenterDepth and CenterDepth* reported in Tables I and IV, the central attribution of gains to the module is wrong.

Watch

Extended reading notes

Core claim

CenterDepth's central claim is that depth estimation in driving scenes does not need a per-pixel depth map; it needs the depth of each object. The system predicts a heatmap of object centers, takes the center as a feature anchor, and uses the Center FC-CRFs module to propagate information from the center to all pixels inside the predicted bounding box, so the object's depth is regressed with global semantic context but without global computation. The paper reports that this design yields a $\delta_1$ of 0.989 and MRE of 0.041 on VirDepth, that it outperforms DepthAnythingV2, MonoDepth2, SMOKE, FCOS3D, DD3D, and BEVFormer on VirDepth, Virtual KITTI 2, and KITTI test sets, and that in the 150-200 m interval its MAE of 3.351 m is far below the 36.269 m of DepthAnythingV2 and 6.354 m of BEVFormer. It further claims the long-range advantage comes specifically from the Center FC-CRFs module, since the same network without it (CenterDepth*) performs markedly worse in the distance-subdivided evaluation.

Load-bearing premise

The load-bearing premise is that the Center FC-CRFs module is a well-defined, trainable component whose described energy minimization is what produces the large accuracy gains, even though the paper leaves its loss and inference procedure unspecified.

Editorial extensions

If this is right

  • If the reported accuracy is real, dense pixel-wise depth maps are not required for obstacle localization: detecting the center and regressing center depth can feed BEV projection and path planning at lower compute cost.
  • Long-range perception beyond 150 m becomes feasible with a single camera, since the reported 150-200 m MAE stays around 3.35 m instead of exploding to tens of meters as it does for full-image estimators.
  • The speed-accuracy trade-off is controllable by backbone choice: ResNet-18 runs at 3.4 ms and Hourglass-104 at 15.6 ms on an RTX 4090, with ResNet-101 near Hourglass accuracy at 9.5 ms.
  • Synthetic datasets generated in hours with KITTI-matched calibration can serve as training and evaluation grounds for depth and 3D detection, reducing the cost of task-specific data collection.
  • Combining the center-point anchor with semantic segmentation and 3D detection objectives in one network can improve both small-object detection (mAP 45.4 vs 30.0 for YOLOv8s in the 150-200 m band) and depth quality.

Reading between the lines

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

  • A direct test of the mechanism would replace Center FC-CRFs with a simple attention-weighted average of features around the center; if the gains persist, the energy formulation itself is not the cause.
  • The reported large-scale balancing is demonstrated inside the synthetic distribution of VirDepth; on real roads, occlusion, weather, and sensor noise at 150-200 m are likely to erode the advantage, so the transfer claim should be read as conditional on the synthetic training distribution.
  • The same center-anchored CRF idea could be applied to depth completion or to anchor-free 3D detectors, where the center heatmap is already computed and the extra cost of a local CRF would be small.
  • Because the paper filters out targets beyond 200 m and highly occluded objects, the method's practical ceiling in denser urban traffic is still unknown; an occlusion-stratified evaluation would be a natural next experiment.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes two contributions: a CARLA-based automated pipeline for generating a synthetic autonomous-driving dataset (VirDepth) with synchronized RGB, depth, semantic, and 3D-box annotations, and a monocular object-centric depth estimation framework called CenterDepth. CenterDepth combines a CenterNet-style keypoint detector with a novel 'Center FC-CRFs' module that is claimed to propagate information from the detected object center within the detection window, yielding high accuracy and long-range stability. The authors report strong results on VirDepth, Virtual KITTI 2, KITTI-Depth, and KITTI-3D, including δ1 of 0.989 and MRE of 0.041 for monocular depth on VirDepth, low MAE in the 150-200 m range, and real-time inference on embedded hardware.

Significance. If the reported results are credible, the paper would offer a low-cost way to generate task-specific virtual driving datasets and a lightweight depth estimator that degrades gracefully at long range. The dataset-generation pipeline is concrete and useful in principle, and the paper explicitly evaluates the difficult 100-200 m regime, which is a genuine practical concern. However, the significance cannot currently be assessed: the central Center FC-CRFs module is not specified at a reproducible level, the ablation baseline is undefined, and the numerical results are internally inconsistent across tables. Because the main technical novelty and the headline accuracy gains both rest on this module, the paper in its present form does not support its central claims.

major comments (4)
  1. [Section III.B, Eqs. (2)-(5)] The Center FC-CRFs module, which is the paper's main technical novelty, is not defined precisely enough to be implemented or evaluated. Eq. (2) defines the center as the arg max of H(x,y), while the text says B(x,y) is the heatmap and H is never introduced. Eq. (5) uses xε,yε without definition. The unary potential ψu is never specified, and the text only says that the final depth is obtained by minimizing E(D); no loss function, network output parametrization, differentiable surrogate, or optimization algorithm is given. Since Tables I, II, and IV attribute the large gap between CenterDepth and CenterDepth* to this module, the central claim is not reproducible from the manuscript.
  2. [Section V.C, Tables I, II, IV] The baseline CenterDepth* is used throughout the experimental section as the comparison point for ablation claims, but it is never defined. The text identifies it only as the 'baseline model' in Section V.C; it is not stated whether it is CenterDepth without the Center FC-CRFs module, without the center anchor, or without some other component, nor how it was trained. Consequently, the claimed improvements attributable to the proposed module, such as δ1 improving from 0.804 to 0.989 on VirDepth with HGL104 and R4 MAE improving from 13.537 to 3.351, cannot be verified or assigned a cause.
  3. [Sections IV, V.A, V.C, V.D and Tables I, III] The manuscript contains multiple contradictory quantitative statements that undermine confidence in the reported results. The dataset is described as containing 'more than 20,000 images' generated in four hours, 'more than 50,000 images' across five scenes, 'roughly 15,000 images' in five scenes in Section V.A, and 'eight urban scenes' in the Introduction. For Hourglass-104, MRE is reported as 0.041 in Table I but as 1.13% in Table III; for ResNet-101, MRE is 0.053 in Table II but 2.31% in Table III. Runtime claims in Section V.D (e.g., 0.025 s for ResNet-18 and 0.038 s for Hourglass-104) do not match Table III values of 3.4 ms and 15.6 ms, respectively. These inconsistencies are too large to be explained by unit conventions.
  4. [Section V.A and V.C] The evaluation protocol for full-image depth baselines is underspecified. The paper does not state whether DepthAnythingV2 and MonoDepth2 were fine-tuned on VirDepth, which checkpoint or input resolution was used, or how their output depth scales were aligned to the metric ground truth before applying the Center or Seg extraction. Because CenterDepth is trained on the same distribution it is tested on, the comparison may mix architectural advantage with training-data exposure, and the reported margins over these baselines cannot be interpreted without this information.
minor comments (5)
  1. [Index Terms] The index term 'Depth Esitimation' is misspelled; it should be 'Depth Estimation'.
  2. [Section III.B] The notation is inconsistent: Eq. (2) uses H(x,y) while the text identifies B(x,y) as the heatmap; please use one symbol consistently and define both the heatmap and the feature map F.
  3. [Section III.B, Eq. (5)] The variables xε,yε in Eq. (5) are undefined; they appear to be intended to denote the center point (xc,yc) used in Eq. (3), but this should be stated explicitly.
  4. [Section V.E] The heading 'Limitations and Systematic Analysis' is misleading: the subsection contains a path-planning demonstration but no discussion of limitations or a systematic error analysis. Please retitle or expand accordingly.
  5. [Section V.E and Table V] Several figure callouts read 'as shown in the figure' without a figure number, and Table V lists Params(M)/FLOPs(B) for CenterDepth that do not match any backbone in Table III; both need clarification.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity identified: the reported depth predictions follow from supervised regression on simulator-generated and public benchmarks, not from the paper's own definitions.

full rationale

The paper's derivation chain is a standard supervised pipeline: VirDepth labels are generated by projecting 3D bounding boxes into camera coordinates (Eqs. 6-8), and CenterDepth is trained to regress target center depth from image features, with the Center FC-CRFs module acting as a regularizer described by an energy function (Eq. 5). The predicted center depth is not defined in terms of the label by construction; it is a network output optimized against ground-truth depth in the usual supervised sense. The pairwise potential in Eq. 3 penalizes deviation from the center depth, but the center depth itself is learned from data, not fitted from the evaluation metric. The comparisons against DepthAnythingV2, MonoDepth2, SMOKE, FCOS3D, DD3D, and BEVFormer on VirDepth, Virtual KITTI 2, and KITTI provide external benchmarks, and KITTI results in Tables I and II give some independent grounding. No load-bearing self-citations or author-imported uniqueness theorems appear; the only methodological citation is to Neural Window FC-CRFs [47] for inspiration, which is not circular. The under-specified equations (undeclared H and xε, missing unary potential and inference procedure) and the internally inconsistent numbers between Tables I, II, and III are serious reproducibility and correctness concerns, but they are not circular reasoning: they do not exhibit a quantity reducing to its own input by definition. Therefore the circularity score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper's central contribution rests on several under-specified choices: the CRF bandwidth and unary potential are not defined, and the transfer from synthetic data to real KITTI is taken for granted. The most significant dependencies are the local window assumption and the feature-similarity-to-depth similarity assumption, both of which enter at Eq. (4) and Eq. (1).

free parameters (3)
  • σ_f (feature similarity bandwidth)
    Used in Eq. 4 to control the pairwise potential weight; no value or tuning procedure is reported, yet it directly controls how strongly pixel depths are pulled toward the center depth.
  • Unary potential ψ_u
    In Eq. 5, ψ_u is never defined; its form (learned or fixed) is a free modeling choice that determines the energy function being minimized.
  • δ threshold (1.10) = 1.10
    The accuracy thresholds δ1, δ2, δ3 use an unusual tolerance of 1.10 (Section V.A). This choice is not justified and inflates the reported δ1 values relative to the standard 1.25 threshold used in the monocular depth literature.
assumptions (5)
  • standard math FC-CRFs can be optimized with tractable approximate inference and integrated into the network
    The paper invokes FC-CRFs (Eq. 5) but never states the inference algorithm (e.g., mean-field) or how gradients flow; the tractability is a background assumption in the cited [47].
  • domain assumption Object depth is well approximated by the center point depth
    The entire CenterDepth design assumes that the center keypoint depth represents the object's distance and that smoothing to that center improves accuracy; this holds only for roughly planar, non-occluded objects.
  • ad hoc to paper Feature similarity in the learned representation correlates with depth similarity
    Eq. 4 weights the pairwise potential by exp(-||f(xi,yi)-f(xc,yc)||^2/(2σ_f^2)), assuming that similar features imply similar depth. No evidence is given that this holds for vehicles or pedestrians at 150-200m.
  • domain assumption CARLA synthetic data transfers to real KITTI without domain adaptation
    The model is trained only on VirDepth but evaluated on KITTI-Depth and KITTI-3D; the reported high KITTI numbers (e.g., δ1=0.963) rest on this transfer assumption.
  • ad hoc to paper The local window (detection bounding box) contains sufficient context for long-range depth
    Center FC-CRFs restricts connections to the bounding box region R (Eq. 1); for tiny distant objects the box may contain few informative pixels, yet the method reports strong R4 performance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Depth as Points: Center Point-based Depth Estimation." pith.science (2026). https://pith.science/paper/U7LNNRTW

@misc{pith2026250418773,
  author       = {Pith},
  title        = {Pith review of: Depth as Points: Center Point-based Depth Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U7LNNRTW}},
  note         = {Machine review of arXiv:2504.18773}
}
read the original abstract

The perception of vehicles and pedestrians in urban scenarios is crucial for autonomous driving. This process typically involves complicated data collection, imposes high computational and hardware demands. To address these limitations, we first develop a highly efficient method for generating virtual datasets, which enables the creation of task- and scenario-specific datasets in a short time. Leveraging this method, we construct the virtual depth estimation dataset VirDepth, a large-scale, multi-task autonomous driving dataset. Subsequently, we propose CenterDepth, a lightweight architecture for monocular depth estimation that ensures high operational efficiency and exhibits superior performance in depth estimation tasks with highly imbalanced height-scale distributions. CenterDepth integrates global semantic information through the innovative Center FC-CRFs algorithm, aggregates multi-scale features based on object key points, and enables detection-based depth estimation of targets. Experiments demonstrate that our proposed method achieves superior performance in terms of both computational speed and prediction accuracy.

Figures

Figures reproduced from arXiv: 2504.18773 by the authors.

Figure 1
Figure 1. Overview of the Proposed System: The architecture consists of an obstacle detection network and a depth prediction module. The CenterDepth module [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The depth prediction results obtained using traditional depth estimation methods are shown. From left to right, each column represents the original [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The left panel illustrates the keypoint prediction of objects using heatmaps, where the intensity of each pixel represents the confidence of the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: A schematic of the virDepth. (a) shows the RGB images, (b) shows the semantic segmentation images, and (c) shows the depth images. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: The training images were selected based on the given semantic [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: The selected results. The first row shows the results on the virDepth, while the first two columns in last row show the results on the Virtual KITTI [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Path planning results based on CenterDepth. As shown in the figure, the first column represents the original images, the second column displays the [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: The figure shows the results of CenterDepth (left) and YOLOv8 (right). Our method demonstrates superior performance in detecting small targets at [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 30 canonical work pages

  1. [1]

    Analyz- ing factors influencing situation awareness in autonomous vehicles—a survey,

    H. A. Ignatious, H. El-Sayed, M. A. Khan, and B. M. Mokhtar, “Analyz- ing factors influencing situation awareness in autonomous vehicles—a survey,” Sensors, vol. 23, no. 8, p. 4075, 2023

  2. [2]

    Applications of computer vision in autonomous vehicles: Methods, challenges and future directions,

    X. Dong and M. L. Cappuccio, “Applications of computer vision in autonomous vehicles: Methods, challenges and future directions,” arXiv preprint arXiv:2311.09093, 2023

  3. [3]

    A review of environmental perception technology based on multi-sensor information fusion in autonomous driving,

    B. Yang, J. Li, and T. Zeng, “A review of environmental perception technology based on multi-sensor information fusion in autonomous driving,” World Electric Vehicle Journal, vol. 16, no. 1, 2025. [Online]. Available: https://www.mdpi.com/2032-6653/16/1/20

  4. [4]

    Recent advances in 3d object detection for self-driving vehicles: A survey,

    O. A. Fawole and D. B. Rawat, “Recent advances in 3d object detection for self-driving vehicles: A survey,” AI, vol. 5, no. 3, pp. 1255–1285,

  5. [5]

    Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving,

    Y . Wang, W.-L. Chao, D. Garg, B. Hariharan, M. Campbell, and K. Q. Weinberger, “Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving,” 2020. [Online]. Available: https://arxiv.org/abs/1812.07179

  6. [6]

    M3d-rpn: Monocular 3d region proposal network for object detection,

    G. Brazil and X. Liu, “M3d-rpn: Monocular 3d region proposal network for object detection,” 2019. [Online]. Available: https: //arxiv.org/abs/1907.06038

  7. [7]

    Ground-aware Monocular 3D Object Detection for Autonomous Driving

    Y . Liu, Y . Yixuan, and M. Liu, “Ground-aware monocular 3d object detection for autonomous driving,” 2021. [Online]. Available: https://arxiv.org/abs/2102.00690

  8. [8]

    Mgnet: Monocular geometric scene understanding for autonomous driving,

    M. Schon, M. Buchholz, and K. Dietmayer, “Mgnet: Monocular geometric scene understanding for autonomous driving,” in 2021 IEEE/CVF International Conference on Computer Vision (ICCV) . IEEE, Oct. 2021, p. 15784–15795. [Online]. Available: http://dx.doi. org/10.1109/ICCV48922.2021.01551

Show all 53 references
  1. [9]

    Vision meets robotics: The kitti dataset,

    A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, “Vision meets robotics: The kitti dataset,” The International Journal of Robotics Research , vol. 32, no. 11, pp. 1231–1237, 2013

  2. [10]

    A2d2: Audi autonomous driving dataset,

    J. Geyer, Y . Kassahun, M. Mahmudi, X. Ricou, R. Durgesh, A. S. Chung, L. Hauswald, V . H. Pham, M. M ¨uhlegg, S. Dorn et al. , “A2d2: Audi autonomous driving dataset,” arXiv preprint arXiv:2004.06320 , 2020

  3. [11]

    Reverse perspec- tive network for perspective-aware object counting,

    Y . Yang, G. Li, Z. Wu, L. Su, Q. Huang, and N. Sebe, “Reverse perspec- tive network for perspective-aware object counting,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 4374–4383

  4. [12]

    Uavid: A semantic segmentation dataset for uav imagery,

    Y . Lyu, G. V osselman, G.-S. Xia, A. Yilmaz, and M. Y . Yang, “Uavid: A semantic segmentation dataset for uav imagery,” ISPRS journal of photogrammetry and remote sensing , vol. 165, pp. 108–119, 2020

  5. [13]

    A fully autonomous indoor quadrotor,

    S. Grzonka, G. Grisetti, and W. Burgard, “A fully autonomous indoor quadrotor,” IEEE Transactions on Robotics , vol. 28, no. 1, pp. 90–100, 2012

  6. [14]

    Vision-based autonomous mapping and exploration using a quadrotor mav,

    F. Fraundorfer, L. Heng, D. Honegger, G. H. Lee, L. Meier, P. Tanskanen, and M. Pollefeys, “Vision-based autonomous mapping and exploration using a quadrotor mav,” in 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems , 2012, pp. 4557–4564

  7. [15]

    Estimation, planning, and mapping for autonomous flight using an rgb-d camera in gps-denied environments,

    A. Bachrach, S. Prentice, R. He, P. Henry, A. S. Huang, M. Krainin, D. Maturana, D. Fox, and N. Roy, “Estimation, planning, and mapping for autonomous flight using an rgb-d camera in gps-denied environments,” The International Journal of Robotics Research, vol. 31, no. 11, pp....

  8. [16]

    nuscenes: A multi- modal dataset for autonomous driving,

    H. Caesar, V . Bankiti, A. H. Lang, S. V ora, V . E. Liong, Q. Xu, A. Krishnan, Y . Pan, G. Baldan, and O. Beijbom, “nuscenes: A multi- modal dataset for autonomous driving,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , June 2020

  9. [17]

    The apolloscape dataset for autonomous driving,

    X. Huang, X. Cheng, Q. Geng, B. Cao, D. Zhou, P. Wang, Y . Lin, and R. Yang, “The apolloscape dataset for autonomous driving,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2018

  10. [18]

    CARLA: An open urban driving simulator,

    A. Dosovitskiy, G. Ros, F. Codevilla, A. Lopez, and V . Koltun, “CARLA: An open urban driving simulator,” in Proceedings of the 1st Annual Conference on Robot Learning , ser. Proceedings of Machine Learning Research, S. Levine, V . Vanhoucke, and K. Goldberg, Eds., vol. 78. PM...

  11. [19]

    Pl-slam: Real-time monocular visual slam with points and lines,

    A. Pumarola, A. Vakhitov, A. Agudo, A. Sanfeliu, and F. Moreno- Noguer, “Pl-slam: Real-time monocular visual slam with points and lines,” in 2017 IEEE international conference on robotics and automa- tion (ICRA). IEEE, 2017, pp. 4503–4508

  12. [20]

    Lift-slam: A deep-learning feature-based monocular visual slam method,

    H. M. S. Bruno and E. L. Colombini, “Lift-slam: A deep-learning feature-based monocular visual slam method,” Neurocomputing, vol. 455, pp. 97–110, 2021

  13. [21]

    Orbeez-slam: A real-time monocular visual slam with orb features and nerf-realized mapping,

    C.-M. Chung, Y .-C. Tseng, Y .-C. Hsu, X.-Q. Shi, Y .-H. Hua, J.-F. Yeh, W.-C. Chen, Y .-T. Chen, and W. H. Hsu, “Orbeez-slam: A real-time monocular visual slam with orb features and nerf-realized mapping,” in 2023 IEEE International Conference on Robotics and Automation (ICRA...

  14. [22]

    Multi-sensor fusion self-supervised deep odometry and depth estimation,

    Y . Wan, Q. Zhao, C. Guo, C. Xu, and L. Fang, “Multi-sensor fusion self-supervised deep odometry and depth estimation,” Remote Sensing, vol. 14, no. 5, p. 1228, 2022

  15. [23]

    Multi-sensor depth fusion framework for real-time 3d reconstruction,

    M. K. Ali, A. Rajput, M. Shahzad, F. Khan, F. Akhtar, and A. B ¨orner, “Multi-sensor depth fusion framework for real-time 3d reconstruction,” Ieee Access, vol. 7, pp. 136 471–136 480, 2019

  16. [24]

    Occfusion: Depth estimation free multi-sensor fusion for 3d occupancy prediction,

    J. Zhang and Y . Ding, “Occfusion: Depth estimation free multi-sensor fusion for 3d occupancy prediction,” arXiv preprint arXiv:2403.05329 , 2024

  17. [25]

    Digging into self-supervised monocular depth estimation,

    C. Godard, O. Mac Aodha, M. Firman, and G. J. Brostow, “Digging into self-supervised monocular depth estimation,” in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 3828– 3838

  18. [26]

    Depth anything: Unleashing the power of large-scale unlabeled data,

    L. Yang, B. Kang, Z. Huang, X. Xu, J. Feng, and H. Zhao, “Depth anything: Unleashing the power of large-scale unlabeled data,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 10 371–10 381

  19. [27]

    Objects as points,

    X. Zhou, D. Wang, and P. Kr ¨ahenb¨uhl, “Objects as points,” arXiv preprint arXiv:1904.07850, 2019

  20. [28]

    Sparse-to-dense: Depth prediction from sparse depth samples and a single image,

    F. Ma and S. Karaman, “Sparse-to-dense: Depth prediction from sparse depth samples and a single image,” in 2018 IEEE international confer- ence on robotics and automation (ICRA) . IEEE, 2018, pp. 4796–4803

  21. [29]

    Learning guided convolutional network for depth completion,

    J. Tang, F.-P. Tian, W. Feng, J. Li, and P. Tan, “Learning guided convolutional network for depth completion,” IEEE Transactions on Image Processing, vol. 30, pp. 1116–1129, 2020

  22. [30]

    Learning joint 2d-3d representations for depth completion,

    Y . Chen, B. Yang, M. Liang, and R. Urtasun, “Learning joint 2d-3d representations for depth completion,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 10 023–10 032

  23. [31]

    Non-local spatial propagation network for depth completion,

    J. Park, K. Joo, Z. Hu, C.-K. Liu, and I. So Kweon, “Non-local spatial propagation network for depth completion,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XIII 16 . Springer, 2020, pp. 120–136

  24. [32]

    Dense surface reconstruction from monocular vision and lidar,

    Z. Li, P. C. Gogia, and M. Kaess, “Dense surface reconstruction from monocular vision and lidar,” in 2019 International Conference on Robotics and Automation (ICRA) . IEEE, 2019, pp. 6905–6911

  25. [33]

    High-precision depth estimation with the 3d lidar and stereo fusion,

    K. Park, S. Kim, and K. Sohn, “High-precision depth estimation with the 3d lidar and stereo fusion,” in 2018 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2018, pp. 2156–2163

  26. [34]

    High-precision depth estimation using uncalibrated lidar and stereo fusion,

    ——, “High-precision depth estimation using uncalibrated lidar and stereo fusion,” Ieee transactions on intelligent transportation systems , vol. 21, no. 1, pp. 321–335, 2019

  27. [35]

    Real-time probabilistic fusion of sparse 3d lidar and dense stereo,

    W. Maddern and P. Newman, “Real-time probabilistic fusion of sparse 3d lidar and dense stereo,” in 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2016, pp. 2181–2188

  28. [36]

    Vpfnet: Improving 3d object detection with virtual point based lidar and stereo data fusion,

    H. Zhu, J. Deng, Y . Zhang, J. Ji, Q. Mao, H. Li, and Y . Zhang, “Vpfnet: Improving 3d object detection with virtual point based lidar and stereo data fusion,” IEEE Transactions on Multimedia, vol. 25, pp. 5291–5304, 2022

  29. [37]

    3d lidar and stereo fusion using stereo matching network with con- ditional cost volume normalization,

    T.-H. Wang, H.-N. Hu, C. H. Lin, Y .-H. Tsai, W.-C. Chiu, and M. Sun, “3d lidar and stereo fusion using stereo matching network with con- ditional cost volume normalization,” in 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2019, pp. 5895–5902

  30. [38]

    3d lidar and stereo fusion using stereo matching network with conditional cost volume normalization. in 2019 ieee,

    ——, “3d lidar and stereo fusion using stereo matching network with conditional cost volume normalization. in 2019 ieee,” in RSJ Interna- tional Conference on Intelligent Robots and Systems (IROS) , pp. 5895– 5902

  31. [39]

    Fcos3d: Fully convolutional one- stage monocular 3d object detection,

    T. Wang, X. Zhu, J. Pang, and D. Lin, “Fcos3d: Fully convolutional one- stage monocular 3d object detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 913–922

  32. [40]

    Smoke: Single-stage monocular 3d object detection via keypoint estimation,

    Z. Liu, Z. Wu, and R. T ´oth, “Smoke: Single-stage monocular 3d object detection via keypoint estimation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 2020, pp. 996–997. 11

  33. [41]

    Depth anything v2,

    L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao, “Depth anything v2,” arXiv preprint arXiv:2406.09414 , 2024

  34. [42]

    Unsupervised monocular depth estimation with left-right consistency,

    C. Godard, O. Mac Aodha, and G. J. Brostow, “Unsupervised monocular depth estimation with left-right consistency,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 270– 279

  35. [43]

    Make3d: Learning 3d scene structure from a single still image,

    A. Saxena, M. Sun, and A. Y . Ng, “Make3d: Learning 3d scene structure from a single still image,” IEEE transactions on pattern analysis and machine intelligence, vol. 31, no. 5, pp. 824–840, 2008

  36. [44]

    A depth estimating method from a single image using foe crf,

    X. Wang, C. Hou, L. Pu, and Y . Hou, “A depth estimating method from a single image using foe crf,” Multimedia Tools and Applications, vol. 74, pp. 9491–9506, 2015

  37. [45]

    Multi-scale contin- uous crfs as sequential deep networks for monocular depth estimation,

    D. Xu, E. Ricci, W. Ouyang, X. Wang, and N. Sebe, “Multi-scale contin- uous crfs as sequential deep networks for monocular depth estimation,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 5354–5362

  38. [46]

    Monocular depth esti- mation using multi-scale continuous crfs as sequential deep networks,

    E. Ricci, W. Ouyang, X. Wang, N. Sebe et al., “Monocular depth esti- mation using multi-scale continuous crfs as sequential deep networks,” IEEE transactions on pattern analysis and machine intelligence, vol. 41, no. 6, pp. 1426–1440, 2018

  39. [47]

    Neural window fully- connected crfs for monocular depth estimation,

    W. Yuan, X. Gu, Z. Dai, S. Zhu, and P. Tan, “Neural window fully- connected crfs for monocular depth estimation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 3916–3925

  40. [48]

    Virtual kitti 2,

    Y . Cabon, N. Murray, and M. Humenberger, “Virtual kitti 2,” arXiv preprint arXiv:2001.10773, 2020

  41. [49]

    Is pseudo- lidar needed for monocular 3d object detection?

    D. Park, R. Ambrus, V . Guizilini, J. Li, and A. Gaidon, “Is pseudo- lidar needed for monocular 3d object detection?” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 3142–3152

  42. [50]

    Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers. arxiv 2022,

    Z. Li, W. Wang, H. Li, E. Xie, C. Sima, T. Lu, Q. Yu, and J. Dai, “Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers. arxiv 2022,” arXiv preprint arXiv:2203.17270

  43. [51]

    Deformable convolutional networks,

    J. Dai, H. Qi, Y . Xiong, Y . Li, G. Zhang, H. Hu, and Y . Wei, “Deformable convolutional networks,” 2017. [Online]. Available: https://arxiv.org/abs/1703.06211

  44. [52]

    An energy and gpu-computation efficient backbone network for real-time object detection,

    Y . Lee, J. won Hwang, S. Lee, Y . Bae, and J. Park, “An energy and gpu-computation efficient backbone network for real-time object detection,” 2019. [Online]. Available: https://arxiv.org/abs/1904.09730

  45. [2024]

    Available: https://www.mdpi.com/2673-2688/5/3/61

    [Online]. Available: https://www.mdpi.com/2673-2688/5/3/61

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.