Pith. sign in

REVIEW 4 major objections 5 minor 40 references

StixelNExT++: Lightweight Monocular Scene Segmentation and Representation for Collective Perception

T0 review · 4 major / 5 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 StixelNExT++ shows that a single monocular RGB frame can be converted end-to-end into a compressed 3D Stixel World at around 10 ms per frame, with performance the authors call competitive within 30 m.

desk verdict A genuinely new 3D Stixel network that ships code and ablations, but the headline 'competitive at 30 m' rests on a bespoke metric that flatters the method. read the letter →

arxiv 2507.06687 v1 pith:JC5EZNLE submitted 2025-07-09 cs.CV cs.RO

Marcel Vosshans , Omar Ait-Aider , Youcef Mezouar , Markus Enzweiler This is my paper
classification cs.CVcs.RO
keywords StixelWorldmonocularscenerepresentationcollectiveperceptionLiDARgroundtruthgenerationdepth-awarelossreal-timeinferenceholisticobstacledetection3Dsegmentation
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

This paper sets out to show that the Stixel World—a compact representation of a scene as vertical sticks anchored in image columns—can be produced end-to-end from a single monocular RGB image, without stereo or temporal context. The network learns to predict, for each depth bin, a Stixel's top, bottom, and probability, and projects the result into metric 3D using known camera intrinsics. Training data are generated automatically from dense LiDAR annotations, including both annotated 3D boxes and a holistic obstacle/free-space rule. The authors report around 10 ms per frame, roughly 28 million parameters, and performance they describe as competitive within a 30 m range, arguing this makes the representation suitable for collective perception where bandwidth and runtime matter.

What carries the argument

The load-bearing object is the Stixel: a vertical stick occupying one image column, defined by top and bottom row and a depth candidate, with a probability of being a real obstacle. Depth is treated as classification over 64 discrete bins from 4 m to 66 m rather than regression, which keeps post-processing to a simple threshold on probability; camera intrinsics then convert each classified Stixel into a 3D point via the inverse projection matrix. Around this, the paper's machinery includes automatically generated LiDAR ground truth (using Patchwork++ for ground/non-ground separation in the holistic case), a depth-aware weighted binary cross-entropy loss that up-weights distant depth candidates, non-linear tangent-based depth discretization that concentrates anchors near the camera, and an optional column cross-attention head that, while not improving average F1, helps at the preferred 30 m operating point. For object-level output, DBSCAN clustering groups adjacent Stixels, with CPU runtimes around 17 ms.

What would settle it

Run the released model on the validation set and compute the official camera-only 3D average precision at a 50% IoU threshold after clustering Stixels into boxes, alongside the reference bounding-box detector; if the clustered Stixels do not reach comparable average precision at 30 m, the claim of competitive 30 m performance fails. A second check: recompute recall under the rule that a ground-truth box is hit only if Stixels cover at least 50% of its projected area, not merely one Stixel inside it.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a lightweight CNN can replace the camera-pair or post-processing machinery traditionally needed for Stixels: one Full-HD RGB frame goes in, and a tensor of 3D Stixels (240 columns, 64 depth candidates, top/bottom/probability) comes out, which can be projected to Cartesian space and clustered into objects. The authors show that the choice of 64 depth candidates with non-linear tangential discretization, a depth-aware weighted binary cross-entropy loss, and a ConvNeXt-Tiny backbone yields the best F1 among their tested configurations, and that the same pipeline can be trained holistically—using LiDAR ground segmentation instead of labeled boxes—to detect obstacles such as buildings, curbs, and unknown objects. They also provide qualitative evidence on KITTI that the holistic variant generalizes to scenes without annotation-based ground planes.

Load-bearing premise

The headline performance claim rests on the paper's own Stixel-specific metric, where a predicted Stixel counts as correct if half its height falls inside a 3D box and a box is recalled if it contains one Stixel; if that metric is rejected, the paper offers no external benchmark for its 30-metre performance.

Editorial extensions

If this is right

  • Because depth is predicted as a small set of column-wise candidates, the output tensor is far smaller than a pixel-wise depth map, making the representation cheap to transmit between vehicles and infrastructure.
  • At roughly 10 ms per frame, the pipeline can keep up with standard camera frame rates on a modern GPU, so the representation can be used as a real-time perception layer rather than an offline analysis tool.
  • The holistic training route means the same network can flag general obstacles—including static structures and unknown objects—without hand-labeled bounding boxes for every class.
  • Clustering Stixels into objects offers a middle ground between 3D bounding boxes and pixel segmentation, preserving shape detail while still supporting object-level reasoning.
  • The approach inherits a dependence on known camera intrinsics: the released weights are tied to the training camera's focal length, and the paper identifies normalizing or predicting focal length as the needed next step.

Reading between the lines

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

  • A direct test of the representation's value would be to plug clustered Stixels into a downstream planner or mapper and measure task performance, such as collision rate or map error, against a bounding-box baseline, since the paper evaluates representation quality rather than end-task benefit.
  • The Stixel evaluation rule (one Stixel inside a box recalls the box) makes recall easy to satisfy; a stricter criterion requiring substantial box coverage by Stixels would show whether the representation truly captures object extent, and would be a natural follow-up metric.
  • The focal-length dependence suggests a concrete extension: retrain with focal length embedded in the input or predicted alongside depth, which the paper already points to, and then test the same weights across different camera rigs.
  • If holistic ground truth from LiDAR ground segmentation is the key to class-agnostic detection, the same recipe could be transferred to other dense-LiDAR datasets, giving a way to benchmark general obstacle detection without manual annotation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper presents StixelNExT++, a neural network that predicts a per-column 3D Stixel representation directly from a single monocular RGB image. The model uses a pretrained ConvNeXt backbone with a lightweight head that, for each of 64 depth candidates per image column, classifies whether a Stixel exists and regresses its top and bottom positions. Depth candidates are discretized over a 4-66 m range. Training data are generated automatically from Waymo LiDAR and annotations using the authors' own Stixel generator. The paper reports ablations over depth-candidate count, backbone choice, a depth-weighted binary cross-entropy loss, non-linear depth discretization, and a column cross-attention module. Performance is evaluated on Waymo with a custom precision/recall metric in which a Stixel is a true positive if 50% of its height lies inside a ground-truth bounding box and a box is recalled if at least one predicted Stixel falls inside it, with PGD as a reference detector. The holistic, annotation-free obstacle-detection variant is shown only qualitatively on KITTI. The abstract claims real-time operation (as low as 10 ms per frame) and competitive performance within a 30-meter range.

Significance. The proposed representation is a plausible middle ground between 3D bounding boxes and dense depth maps, and the open release of the model, the Stixel generator, and the pyStixel-lib devkit is a useful community resource. The ablation study provides some evidence about design choices such as depth-bin count, loss weighting, and backbone. However, the central quantitative claim of 'competitive performance' rests on a metric invented for this paper and on a baseline that the authors themselves disavow. If the evaluation were supplemented with established Stixel benchmarks (e.g., comparisons with Mono-Stixel or StixelNet), error bars, and a precise runtime protocol, the contribution would be significant for lightweight collective perception. As it stands, the experimental validation is not yet at the level needed to support the abstract's headline claims.

major comments (4)
  1. [III-B] The evaluation uses a metric defined in this paper rather than an established Stixel benchmark. A predicted Stixel is accepted if at least 50% of its height lies inside a 3D bounding box, and a ground-truth box is recalled if a single predicted Stixel falls inside it. Because the network outputs up to 64 depth candidates per column, a single spurious candidate at the correct depth can count as a hit, making recall easy to inflate, and precision does not measure geometric or depth accuracy. No error bars are provided and no external Stixel method is compared under the same protocol. This makes the abstract's 'competitive performance within a 30-meter range' unsupported by the current evidence.
  2. [III-B, Table II] The only quantitative baseline, PGD, is a 3D bounding box detector evaluated with an IoU threshold of 50%, which the paper notes is 'slightly more lenient than the official criteria,' and the text immediately states that the comparison 'is not intended as a direct baseline.' The F1 comparison in Table II (StixelNExT++ 60-64% vs. PGD 35.6%) is therefore not a legitimate competitive comparison. The paper also does not implement or report any prior Stixel method (Mono-Stixel [9], StixelNet [7]), so there is no evidence that the learned representation outperforms or even matches existing Stixel approaches.
  3. [IV-C] The claimed holistic obstacle detection is supported only by qualitative examples on KITTI (Fig. 7); no quantitative metric is reported for the holistic variant. Since the abstract lists generalized obstacle detection via a holistic approach as a contribution, the absence of a quantitative evaluation of this mode is a gap that should be closed, for example by measuring free-space or obstacle IoU against the LiDAR ground truth used in training.
  4. [IV-B, Table III] The abstract's 'computation times as low as 10 ms per frame' is not backed by the runtime table: Table III reports GPU passthrough times on random tensors (e.g., 1.44 ms for the ConvNeXt backbone), and the 10 ms figure does not appear in the experiments. The runtime measurement uses random tensors rather than real images and does not report end-to-end latency including the full post-processing and clustering steps. The real-time claim should be either substantiated with a precise measurement protocol or removed.
minor comments (5)
  1. [Table III] In Table III, '22, 2 M' should read '22.2 M'; the table also reports single runtime values without standard deviations, which would be helpful given the 1,000-sample measurement.
  2. [Eq. (9)] The parameter 'a' in the tangent-based depth discretization is introduced but no value or selection criterion is reported, despite being a free parameter of the method.
  3. [Fig. 4] Figure 4's caption cites F1-Scores and segmentation IoU values (e.g., 35%/67%) that are not defined in the text; please clarify which metrics these are and how they relate to Section III-B.
  4. [Author biographies] In the author bio, 'Autonmous' should be 'Autonomous.'
  5. [III-A] The phrase 'passthrough size of ≈ 6.4 GB' is undefined; please clarify whether this refers to the GPU memory footprint of a full-resolution forward pass, as the number is not otherwise justified.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the system is an empirical model whose headline claim is evaluated against external Waymo data, though the custom Stixel metric is lenient.

full rationale

No circular derivation chain exists in this paper. StixelNExT++ is an empirical learning system: it is trained on Stixel ground truth generated from Waymo LiDAR and annotations using the authors' earlier generator [3], and the training loss (Eq. 7) supervises Stixel properties against that ground truth. The headline claim of 'competitive performance within a 30-meter range' is then measured against external Waymo 3D bounding boxes using a purpose-built Stixel metric (Sec. III-B). That metric is lenient, and the only reference baseline, PGD, is compared with a more permissive IoU threshold, which weakens the strength of the claim. However, leniency is not circularity: a predicted Stixel can still fail the metric, the metric is not a term in the training loss, and the evaluation is not defined in terms of the network's own outputs. The self-citations to [3] and [2] provide the data-generation tooling and the AEIF dataset, but the central performance claim is not proved by those citations; it rests on external Waymo evaluations, with the paper itself acknowledging in Sec. V that adapting box metrics to Stixels constitutes a 'slightly ill-posed comparison.' No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' previous work, and no derivation reduces by construction to its inputs. Therefore the appropriate finding is no significant circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on an empirical pipeline: a neural network trained on LiDAR-derived Stixel GT and evaluated with custom metrics. The free parameters are architecture and loss choices, several tuned on validation data; none is derived from first principles. The axioms are standard camera geometry plus domain assumptions about the fidelity of the LiDAR GT and the ground segmentation. No invented entities are introduced.

free parameters (6)
  • depth_candidates = 64
    Ablation across 32, 64, 128, 192 depth anchors; 64 selected because it gave the highest F1 in Table II. This is an architecture hyperparameter tuned on validation data.
  • stixel_width = 8 pixels
    Output columns are 240 for a 1920-wide image, i.e., 8-pixel columns; used consistently for GT generation, network output, and IoU sampling.
  • depth_range = 4 to 66 m
    Depth candidates discretize 4-66 m in Eq. (9); the range is a design choice, not derived in the paper.
  • wbc_alpha_scale = alpha_min=1x, alpha_max=2x
    Depth-aware WBCE loss uses a linear multiplier from 1x to 2x over the depth range (Eq. 8); values are hand-set.
  • tangent_limiter_a = not reported
    Non-linear depth discretization Eq. (9) depends on a limiting factor 'a' for the tangent function; the numerical value is not given in the paper.
  • operating_threshold = 38%
    The preferred operating point for the 30 m claim uses a probability threshold of 38%; chosen from the PR curve, not from a held-out criterion.
assumptions (4)
  • standard math Pinhole camera projection with known intrinsics (Eqs. 2-6)
    The 3D reconstruction uses the standard Hartley-Zisserman projection model; assumed valid for the Waymo front camera.
  • domain assumption LiDAR-based Stixel ground truth is a faithful supervision signal
    Section III-A: the automatic GT generator [3] converts LiDAR and semantic annotations into Stixels; if this conversion is inaccurate, the network learns the generator's errors.
  • domain assumption Patchwork++ ground segmentation correctly separates ground and non-ground points for holistic training
    Section IV-C: the holistic approach relies on Patchwork++ rather than explicit object ground planes; the authors note inaccuracies on hilly or uneven terrain.
  • domain assumption Waymo ground-truth boxes and semantic labels are accurate enough for evaluation
    Used in relevance filtering and IoU evaluation; no annotation-quality analysis is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of StixelNExT++: Lightweight Monocular Scene Segmentation and Representation for Collective Perception." pith.science (2026). https://pith.science/paper/JC5EZNLE

@misc{pith2026250706687,
  author       = {Pith},
  title        = {Pith review of: StixelNExT++: Lightweight Monocular Scene Segmentation and Representation for Collective Perception},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JC5EZNLE}},
  note         = {Machine review of arXiv:2507.06687}
}
read the original abstract

This paper presents StixelNExT++, a novel approach to scene representation for monocular perception systems. Building on the established Stixel representation, our method infers 3D Stixels and enhances object segmentation by clustering smaller 3D Stixel units. The approach achieves high compression of scene information while remaining adaptable to point cloud and bird's-eye-view representations. Our lightweight neural network, trained on automatically generated LiDAR-based ground truth, achieves real-time performance with computation times as low as 10 ms per frame. Experimental results on the Waymo dataset demonstrate competitive performance within a 30-meter range, highlighting the potential of StixelNExT++ for collective perception in autonomous systems.

Figures

Figures reproduced from arXiv: 2507.06687 by the authors.

Figure 1
Figure 1. A StixelNExT++ result is depicted on the left in the 2D Stixel World Space, where relative depth is color-coded from red (near) to green (far). On the right, the corresponding prediction is visualized in 3D space with absolute metric depth, while maintaining the relative RGB values for enhanced comparison. StixelNExT++ neural network generates a 3D scene representation directly from monocular RGB images in an end￾to… view at source ↗
Figure 2
Figure 2. The Stixel World representation spectrum illustrates different levels of scene abstraction for robotics perception tasks. At the coarse end, 3D bounding boxes offer compact, object-level abstractions optimized for cluster-based detection. At the fine end, pixel-wise depth or occupancy maps provide detailed, high-resolution scene representations. Stixel World adaptations occupy an intermediate position, balancing det… view at source ↗
Figure 3
Figure 3. The concept of StixelNExT++ involves processing a single monocular RGB image, directly at full resolution, using a pretrained ConvNeXt [27] backbone. A lightweight head is attached to shape the extracted features and interpret the predictions. The network is designed to directly learn a Stixel representation, leveraging known camera intrinsics to project the Stixels into 3D space. Additionally, we investigated a col… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Sample (a-c): Qualitative inference results. Colors indicate predicted relative depth in the Stixel predictions. Yellow arrows show the field of view [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: , performance decreased for probabilities above 70%. However, it performed best at our preferred operating point (up to 30 meters with a threshold of 38%), achieving a recall of 89.3% (compared to 89.2% for the second-best model) and a precision of 88.3% (compared to 8…
Figure 6
Figure 6. Figure 6: Two clustering results of a predicted Stixel World using the DBSCAN algorithm. Both examples showed promising outcomes with CPU runtimes of 16.5 ms (left) and 17.1 ms (right), respectively. hardware to facilitate a direct comparison. C. Qualitative Results Clustering. …
Figure 7
Figure 7. Figure 7: A qualitative result from holistic training on the KITTI dataset [31] is presented. Each sample illustrates, from top to bottom: the input image, the predicted depth Stixels, and the clustered Stixels. our method adapted the representation for monocular inputs, extendi…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 25 canonical work pages

  1. [9]

    Mono-Stixels: Monocular depth reconstruction of dynamic street scenes

    F. Brickwedde, S. Abraham, and R. Mester, “Mono-Stixels: Monocular depth reconstruction of dynamic street scenes,” in 2018 IEEE International Conference on Robotics and Automation (ICRA) , 2019, pp. 3369–3375. [Online]. Available: https://arxiv.org/abs/1908.02635

    work page Pith review arXiv 2018

  2. [7]

    StixelNet: A Deep Convolutional Network for Obstacle Detection and Road Segmentation,

    D. Levi, N. Garnett, and E. Fetaya, “StixelNet: A Deep Convolutional Network for Obstacle Detection and Road Segmentation,” in British Machine Vision Conference 2015 . Swansea: British Machine Vision Association, 2015, pp. 109.1–109.12. [Online]. Available: http: //www.bmva.org/bmvc/2015/papers/paper109/index.html

    work page 2015

  3. [1]

    The Stixel World - A Compact Medium Level Representation of the 3D-World,

    H. Badino, U. Franke, and D. Pfeiffer, “The Stixel World - A Compact Medium Level Representation of the 3D-World,” in Pattern Recognition, J. Denzler, G. Notni, and H. S ¨uße, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 2009, vol. 5748, pp. 51–60. [Online]. Available: http://link.springer.com/10.1007/978-3-642-03798-6 \ 6

    work page doi:10.1007/978-3-642-03798-6 2009

  4. [2]

    The AEIF Data Collection: A Dataset for Infrastructure-Supported Perception Research with Focus on Public Transportation,

    M. V osshans, A. Baumann, M. Drueppel et al. , “The AEIF Data Collection: A Dataset for Infrastructure-Supported Perception Research with Focus on Public Transportation,” 2024, version Number: 2. [Online]. Available: https://arxiv.org/abs/2407.08261

    work page arXiv 2024

  5. [3]

    StixelNExT: Toward Monocular Low-Weight Perception for Object Segmentation and Free Space Detection,

    M. V osshans, O. Ait-Aider, Y . Mezouar, and M. Enzweiler, “StixelNExT: Toward Monocular Low-Weight Perception for Object Segmentation and Free Space Detection,” in 2024 IEEE Intelligent Vehicles Symposium (IV). Jeju Island, Korea, Republic of: IEEE, Jun. 2024, pp. 2154–2161. [Online]. Available: https://ieeexplore.ieee.org/document/10588680/

    work page arXiv 2024

  6. [4]

    Towards a Global Optimal Multi-Layer Stixel Representation of Dense 3D Data,

    D. Pfeiffer and U. Franke, “Towards a Global Optimal Multi-Layer Stixel Representation of Dense 3D Data,” in British Machine Vision Conference 2011. Dundee: British Machine Vision Association, 2011, pp. 51.1–51.12. [Online]. Available: http://www.bmva.org/bmvc/2011/ proceedings/paper51/index.html

    work page 2011

  7. [5]

    Semantic Stixels: Depth is not enough,

    L. Schneider, M. Cordts, T. Rehfeld et al. , “Semantic Stixels: Depth is not enough,” in 2016 IEEE Intelligent Vehicles Symposium (IV) . Gotenburg, Sweden: IEEE, Jun. 2016, pp. 110–117. [Online]. Available: http://ieeexplore.ieee.org/document/7535373/

    work page arXiv 2016

  8. [6]

    Instance Stixels: Segmenting and Grouping Stixels into Objects,

    T. M. Hehn, J. F. P. Kooij, and D. M. Gavrila, “Instance Stixels: Segmenting and Grouping Stixels into Objects,” in 2019 IEEE Intelligent Vehicles Symposium (IV) . Paris, France: IEEE, Jun. 2019, pp. 2542–2549. [Online]. Available: https://ieeexplore.ieee.org/ document/8814243/

    work page arXiv 2019

Show all 40 references
  1. [8]

    Real-Time Category- Based and General Obstacle Detection for Autonomous Driving,

    N. Garnett, S. Silberstein, S. Oron et al. , “Real-Time Category- Based and General Obstacle Detection for Autonomous Driving,” in 2017 IEEE International Conference on Computer Vision Workshop (ICCVW). Venice: IEEE, Oct. 2017, pp. 198–205. [Online]. Available: http://ieeexplo...

    2017

  2. [10]

    Exploiting Single Image Depth Prediction for Mono-stixel Estimation,

    Brickwedde, Abraham, and Mester, “Exploiting Single Image Depth Prediction for Mono-stixel Estimation,” in Computer Vision – ECCV 2018 Workshops , 2019. [Online]. Available: https://link.springer.com/ 10.1007/978-3-030-11009-3 \ 14

    2018 doi

  3. [11]

    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 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) ,

    2017

  4. [12]

    Digging Into Self-Supervised Monocular Depth Estimation,

    C. Godard, O. Mac Aodha, M. Firman, and G. Brostow, “Digging Into Self-Supervised Monocular Depth Estimation,” in Proceedings of the IEEE/CVF international conference on computer vision , 2018, pp. 3828–3838. [Online]. Available: https://arxiv.org/abs/1806.01260

    2018 arXiv

  5. [13]

    Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data,

    L. Yang, B. Kang, Z. Huang et al. , “Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data,” Apr. 2024, arXiv:2401.10891 [cs]. [Online]. Available: http://arxiv.org/abs/2401.10891

    2024 arXiv

  6. [14]

    Depth Anything V2,

    L. Yang, B. Kang, Z. Huang, and Z. Zhao, “Depth Anything V2,”

  7. [15]

    MonoScene: Monocular 3D Semantic Scene Completion,

    A.-Q. Cao and R. de Charette, “MonoScene: Monocular 3D Semantic Scene Completion,” in 2022 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2021. [Online]. Available: https://arxiv.org/abs/2112.00726

    2022 arXiv

  8. [16]

    Learning Occupancy for Monocular 3D Object Detection,

    L. Peng, J. Xu, H. Cheng et al. , “Learning Occupancy for Monocular 3D Object Detection,” May 2023, arXiv:2305.15694 [cs]. [Online]. Available: http://arxiv.org/abs/2305.15694

    2023 arXiv

  9. [17]

    MonoGRNet: A Geometric Reasoning Network for Monocular 3D Object Localization,

    Z. Qin, J. Wang, and Y . Lu, “MonoGRNet: A Geometric Reasoning Network for Monocular 3D Object Localization,” Mar. 2020, arXiv:1811.10247 [cs]. [Online]. Available: http://arxiv.org/abs/1811. 10247

    2020 arXiv

  10. [18]

    You Only Look Bottom-Up for Monocular 3D Object Detection,

    K. Xiong, D. Zhang, D. Liang et al. , “You Only Look Bottom-Up for Monocular 3D Object Detection,” IEEE Robotics and Automation Letters, vol. 8, no. 11, pp. 7464–7471, Nov. 2023, arXiv:2401.15319 [cs]. [Online]. Available: http://arxiv.org/abs/2401.15319

    2023 arXiv

  11. [19]

    Translating Images into Maps,

    A. Saha, O. M. Maldonado, C. Russell, and R. Bowden, “Translating Images into Maps,” Mar. 2022, arXiv:2110.00966 [cs]. [Online]. Available: http://arxiv.org/abs/2110.00966

    2022 arXiv

  12. [20]

    SeaBird: Segmentation in Bird’s View with Dice Loss Improves Monocular 3D Detection of Large Objects,

    A. Kumar, Y . Guo, X. Huang et al. , “SeaBird: Segmentation in Bird’s View with Dice Loss Improves Monocular 3D Detection of Large Objects,” in Conference on Computer Vision and Pattern Recognition (CVPR) 2024, 2024

    2024

  13. [21]

    Enhancing 3D Object Detection with 2D Detection-Guided Query Anchors,

    H. Ji, P. Liang, and E. Cheng, “Enhancing 3D Object Detection with 2D Detection-Guided Query Anchors,” 2024, version Number: 1. [Online]. Available: https://arxiv.org/abs/2403.06093

    2024 arXiv

  14. [22]

    Estimating Depth From Monocular Images as Classification Using Deep Fully Convolutional Residual Networks,

    Y . Cao, Z. Wu, and C. Shen, “Estimating Depth From Monocular Images as Classification Using Deep Fully Convolutional Residual Networks,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 28, no. 11, pp. 3174–3182, Nov. 2018. [Online]. Available: https://iee...

    2018

  15. [23]

    Pseudo-LiDAR From Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving,

    Y . Wang, W.-L. Chao, D. Garg et al. , “Pseudo-LiDAR From Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving,” in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . Long Beach, CA, USA: IEEE, Jun. 2019, pp. 8437–84...

    2019

  16. [24]

    CAM-Convs: Camera- Aware Multi-Scale Convolutions for Single-View Depth,

    J. M. Facil, B. Ummenhofer, H. Zhou et al. , “CAM-Convs: Camera- Aware Multi-Scale Convolutions for Single-View Depth,” Apr. 2019, arXiv:1904.02028. [Online]. Available: http://arxiv.org/abs/1904.02028

    2019 arXiv

  17. [25]

    Learning Depth from Single Images with Deep Neural Network Embedding Focal Length,

    L. He, G. Wang, and Z. Hu, “Learning Depth from Single Images with Deep Neural Network Embedding Focal Length,” Mar. 2018, arXiv:1803.10039. [Online]. Available: http://arxiv.org/abs/1803.10039

    2018 arXiv

  18. [26]

    Patchwork++: Fast and Robust Ground Segmentation Solving Partial Under-Segmentation Using 3D Point Cloud,

    S. Lee, H. Lim, and H. Myung, “Patchwork++: Fast and Robust Ground Segmentation Solving Partial Under-Segmentation Using 3D Point Cloud,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . Kyoto, Japan: IEEE, Oct. 2022, pp. 13 276–13 283. [Onl...

    2022

  19. [27]

    A ConvNet for the 2020s,

    Z. Liu, H. Mao, C.-Y . Wu et al. , “A ConvNet for the 2020s,” in Conference on Computer Vision and Pattern Recognition (CVPR) , 2022, publisher: arXiv Version Number: 2. [Online]. Available: https://arxiv.org/abs/2201.03545

    2022 arXiv

  20. [28]

    Hartley and A

    R. Hartley and A. Zisserman, Multiple View Geometry in Computer Vision , 2nd ed. Cambridge University Press, Mar. 2004. [Online]. Available: https://www.cambridge.org/core/product/identifier/ 9780511811685/type/book

    2004

  21. [29]

    Scalability in Perception for Autonomous Driving: Waymo Open Dataset,

    P. Sun, H. Kretzschmar, X. Dotiwalla et al., “Scalability in Perception for Autonomous Driving: Waymo Open Dataset,” in Conference on Computer Vision and Pattern Recognition (CVPR) 2020, 2019. [Online]. Available: https://arxiv.org/abs/1912.04838

    2020 arXiv

  22. [30]

    LET-3D-AP: Longitudinal Error Tolerant 3D Average Precision for Camera- Only 3D Detection,

    W.-C. Hung, V . Casser, H. Kretzschmar et al. , “LET-3D-AP: Longitudinal Error Tolerant 3D Average Precision for Camera- Only 3D Detection,” 2022, version Number: 2. [Online]. Available: https://arxiv.org/abs/2206.07705

    2022 arXiv

  23. [31]

    Are we ready for autonomous driving? The KITTI vision benchmark suite,

    A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? The KITTI vision benchmark suite,” in 2012 IEEE Conference on Computer Vision and Pattern Recognition . Providence, RI: IEEE, Jun. 2012, pp. 3354–3361. [Online]. Available: http: //ieeexplore.ieee.org/do...

    2012

  24. [32]

    Probabilistic and Geometric Depth: Detecting Objects in Perspective,

    T. Wang, X. Zhu, J. Pang, and D. Lin, “Probabilistic and Geometric Depth: Detecting Objects in Perspective,” 2021, version Number: 3. [Online]. Available: https://arxiv.org/abs/2107.14160

    2021 arXiv

  25. [33]

    The Pascal Visual Object Classes Challenge: A Retrospective,

    M. Everingham, S. M. A. Eslami, L. Van Gool et al., “The Pascal Visual Object Classes Challenge: A Retrospective,” International Journal of Computer Vision , vol. 111, no. 1, pp. 98–136, Jan. 2015. [Online]. Available: http://link.springer.com/10.1007/s11263-014-0733-5

    2015 doi

  26. [34]

    EfficientNetV2: Smaller Models and Faster Training,

    M. Tan and Q. V . Le, “EfficientNetV2: Smaller Models and Faster Training,” in 2021 International Conference on Machine Learning , 2021, publisher: arXiv Version Number: 3. [Online]. Available: https://arxiv.org/abs/2104.00298

    2021 arXiv

  27. [35]

    Searching for MobileNetV3,

    A. Howard, M. Sandler, G. Chu et al. , “Searching for MobileNetV3,” 2019, version Number: 5. [Online]. Available: https://arxiv.org/abs/ 1905.02244

    2019 arXiv

  28. [36]

    ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design,

    N. Ma, X. Zhang, H.-T. Zheng, and J. Sun, “ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design,” 2018, version Number: 1. [Online]. Available: https://arxiv.org/abs/1807.11164

    2018 arXiv

  29. [37]

    Swin Transformer: Hierarchical Vision Transformer using Shifted Windows,

    Z. Liu, Y . Lin, Y . Cao et al. , “Swin Transformer: Hierarchical Vision Transformer using Shifted Windows,” in 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021, publisher: arXiv Version Number: 2. [Online]. Available: https://arxiv.org/abs/2103.14030

    2021 arXiv

  30. [38]

    Attention Is All You Need,

    A. Vaswani, N. Shazeer, N. Parmar et al. , “Attention Is All You Need,” Advances in neural information processing systems (NIPS) , vol. 30, 2017, publisher: arXiv Version Number: 7. [Online]. Available: https://arxiv.org/abs/1706.03762 Marcel V osshans received his Bachelor’s ...

    2017 arXiv

  31. [2016]

    Available: https://arxiv.org/abs/1609.03677

    [Online]. Available: https://arxiv.org/abs/1609.03677

    arXiv

  32. [2024]

    Available: http://arxiv.org/abs/2406.09414

    [Online]. Available: http://arxiv.org/abs/2406.09414

    arXiv

Pith tools

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