Pith. sign in

REVIEW 5 major objections 6 minor 54 references

Rethink 3D Object Detection from Physical World

T0 review · 5 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 The paper's new latency- and planning-aware metrics rank 3D detectors by real-time usefulness, not just accuracy.

desk verdict A practical pair of metrics, L-AP and P-AP, that deserve engagement, but the empirical validation is built on injected synthetic errors and same-split HPO, so the benchmark claims should be treated as provisional. read the letter →

arxiv 2507.00190 v1 pith:3D63NNIR submitted 2025-06-30 cs.RO cs.CV

Satoshi Tanaka , Koji Minoda , Fumiya Watanabe , Takamasa Horibe This is my paper
classification cs.ROcs.CV
keywords 3Dobjectdetectionautonomousdrivingevaluationmetricslatency-awareAPplanning-awarehyperparameteroptimizationLiDARpointcloudmotionplanning
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 mean average precision (mAP), the standard score for 3D object detection, ignores two physical facts that matter in autonomous driving: inference takes time, and only some detection errors endanger the vehicle. To fix this, it introduces latency-aware AP (L-AP), which measures accuracy at the moment the inference actually completes, and planning-aware AP (P-AP), which filters out occluded objects and penalizes only the misdetections that can cause a collision. Using the nuPlan closed-loop simulator, it shows that when artificial latency, yaw, and position errors are injected into ground truth, L-AP and P-AP fall in step with the planner's safety score while mAP stays flat. It then tunes a CenterPoint-based model with L-AP as the objective (L-HPO), reaching an L-mAP of 55.1 on an RTX3090, and uses the metric to show that adding more LiDAR frames can hurt rather than help real-time performance.

What carries the argument

The central objects are the two metrics. L-AP evaluates detection at time $t+\Delta t$, where $\Delta t$ is the inference latency, by estimating the ground truth position as $x^{GT}_{t+\Delta t} = x^{GT}_t + \hat{v}^{GT}_t \Delta t$ and the predicted position as $x^{Pred}_{t+\Delta t} = x^{Pred}_t + v^{Pred}_t \Delta t - v^{Ego}_t \Delta t$, so a slow detector is penalized by how far objects move while it thinks. P-AP restricts scoring to non-occluded planning-relevant objects, matches on the distance to the nearest surface instead of the center with a planning margin $d_m = 0.5$ m, and computes the average distance of the four bounding-box corners, so that yaw errors on large vehicles hurt more than on small ones and 'mis-detection as farther away' counts as unsafe. Together they give a single evaluation axis on which latency, hardware, accelerators, and model architecture can be compared.

What would settle it

Take a set of real detectors on nuScenes, compute L-AP and P-AP for each, then run each detector's outputs through the nuPlan closed-loop planner and rank by nuPlan score; if the L-AP/P-AP ranking disagrees materially with the planning ranking, or if two detectors with equal L-AP/P-AP produce clearly different nuPlan scores, the central claim fails.

Watch

Extended reading notes

Core claim

The central claim is that the trade-off between speed and accuracy, and the difference between safe and unsafe misdetections, can be captured by two single-axis metrics: L-AP and P-AP. L-AP shifts each detection forward in time by the inference latency, comparing the predicted object to where the ground truth will be when the result becomes available; P-AP restricts evaluation to objects that matter for motion planning, replacing center distance with mean corner displacement and treating overestimates of distance beyond a margin as mismatches. On nuPlan, both metrics track the closed-loop score as injected errors grow, while mAP remains near perfect, showing that the new metrics reflect planning quality that mAP misses. Building on this, the paper's L-HPO model, a lightweight CenterPoint variant optimized along the L-AP axis, achieves 55.1 L-mAP on an RTX3090 with PyTorch, and the same axis makes hardware, accelerator, and model choices comparable.

Load-bearing premise

The validation that L-AP and P-AP track planning quality is done by injecting artificial latency, yaw, and position errors into ground-truth detections, then reading the nuPlan score; this assumes such synthetic errors resemble the real, possibly correlated errors of actual detectors closely enough that the correlation carries over.

Editorial extensions

If this is right

  • Detector rankings change: a model with slightly higher mAP but much higher latency (e.g., BEVFusion vs TransFusion-L) loses to the faster model in L-mAP, reversing the usual leaderboard.
  • Hardware and accelerator choices become comparable on one axis: the paper shows that 'RTX4060Ti + TensorRT + CenterPoint' outscores 'RTX3090 + PyTorch + TransFusion-L' in L-mAP, which plain mAP cannot express.
  • Adding more LiDAR frames is not always beneficial: merging 9 frames raises mAP to 56.3 but lowers L-mAP to 43.7, so point-cloud density must be balanced against the latency it adds.
  • Optimization can be done directly along the metric axis: L-HPO, a CenterPoint variant tuned for L-mAP, reaches 55.1 L-mAP on an RTX3090 with PyTorch, outperforming heavier fusion models in real-time use.

Reading between the lines

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

  • The same axis could extend beyond detection to tracking and motion prediction, which the paper names as future work; the latency-shift idea applies wherever a downstream planner consumes stale estimates.
  • One could test the metric on real, correlated detector failures: the current validation injects errors per-object independently, so a natural extension is to corrupt whole scenes with realistic failure modes (e.g., missing a specific object class at a distance) and check whether P-AP still tracks the closed-loop score.
  • Because L-AP folds hardware latency into the score, it effectively converts a deployment problem (which GPU or accelerator) into a pure ranking problem; this could be used to auto-select hardware-software configurations at deploy time, e.g., through a cost-table lookup of the kind the paper sketches in its cost study.
  • The corner-distance matching in P-AP is one concrete choice for 'planning-aware' geometry; alternative distances (e.g., time-to-collision weighted) might yield even tighter alignment with specific planners, and the paper's framework allows that substitution.
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

5 major / 6 minor

Summary. The paper proposes two new evaluation metrics for real-time 3D object detection: Latency-aware AP (L-AP), which shifts predicted and ground-truth boxes by velocity times inference latency before computing AP, and Planning-aware AP (P-AP), which applies occlusion filtering, a planning margin on the distance to the nearest surface, and corner-distance matching to focus on objects relevant to motion planning. The authors validate these metrics on the nuPlan benchmark by injecting latency, yaw, and position errors into ground-truth detections and comparing L-mAP/P-mAP against the nuPlan closed-loop score. They then benchmark several LiDAR, camera, and fusion detectors on nuScenes, introduce L-HPO (a CenterPoint variant tuned for L-mAP), and use L-mAP to analyze hardware and accelerator trade-offs, multi-frame point-cloud fusion, and development-cost optimization. The central claim is that L-AP and P-AP provide a single-axis, planning-relevant evaluation that better reflects real-time deployment than mAP plus latency separately.

Significance. If the central claim is established, the proposed metrics would be a useful addition to the evaluation toolbox for real-time autonomous driving perception, particularly for practitioners who must choose among hardware, accelerators, and model architectures. The cost analysis in Sec. 4.5 is a concrete, practical illustration of how a single-axis latency-aware metric can support deployment decisions. The paper is also commendable for making explicit the distinction between safe and unsafe misdetections and for quantitatively showing that merging more LiDAR frames is not always beneficial in real-time settings. However, the current evidence is insufficient to support the claim that L-AP and P-AP rank real detectors by planning quality: the validation uses synthetic perturbations of ground truth rather than real detector outputs, the P-AP margin dm is hand-chosen, the baseline CenterPoint configuration is inconsistent across tables, and L-HPO is tuned and evaluated on the same validation split.

major comments (5)
  1. [Sec. 4.2, Table 3] The metric validation replaces detector outputs with ground-truth boxes perturbed by uniform latency, yaw, and position errors. Real 3D detector errors are spatially and temporally correlated (e.g., missed or shifted clusters of objects, range-dependent depth bias, class-specific failure modes), so the demonstrated correlation between L-mAP/P-mAP and nuPlan closed-loop score may not transfer to actual detectors. The central claim that these metrics reflect planning quality for real detectors requires at least one experiment with actual detector outputs (e.g., the models in Table 4) evaluated in the nuPlan closed loop, or an error-injection procedure calibrated to the error distributions of those detectors.
  2. [Sec. 4.3 vs Sec. 4.4, Tables 4 and 6] The CenterPoint baseline is reported inconsistently. Table 4 gives latency 80 ms, mAP 56.3, L-mAP 46.7, and P-mAP 36.9, while Table 6, N=1 row gives the same model latency 31 ms, mAP 46.7, L-mAP 45.5, and P-mAP 28.5. These numbers cannot describe the same configuration, and Table 6 never reproduces the Table 4 values for any N. Since the claimed +8.4 L-mAP improvement of L-HPO is measured against this baseline, the baseline must be defined consistently and the discrepancy reconciled.
  3. [Sec. 4.2, Table 3(c)] For objects closer than 20 m shifted by -0.6 m, P-mAP decreases from 100.0 to 76.6 while the nuPlan score increases from 0.9277 to 0.9354. The paper acknowledges the nuPlan improvement but still interprets P-mAP as tracking planning quality. This is a direct counterexample to monotonicity and shows that P-mAP penalizes safe closer misdetections. The paper should either justify this behavior as desirable or modify P-mAP so that closer-than-GT errors within dm do not reduce the score; in either case a sensitivity analysis with respect to dm=0.5 is needed.
  4. [Secs. 3.1 and 3.2] L-AP and P-AP are described in prose and examples, but no complete mathematical definition is given. For P-AP, it is unclear how the occlusion filter, the de>dm matching rule, and the corner-distance thresholds are combined into a precision-recall curve and a mean AP across classes and distance thresholds. For L-AP, the transformation of predicted and GT boxes is specified in Eqs. (1)-(3), but the subsequent AP computation is not. A metrics paper must provide a reproducible algorithmic definition, including the exact threshold sets and matching procedure, otherwise the numerical results cannot be verified.
  5. [Sec. 4.3, L-HPO] L-HPO is optimized on the nuScenes validation split and reported on the same split, and the search space, budget, and number of trials are not given. The +8.4 L-mAP gain over the default CenterPoint may reflect tuning of multi-frame merging, voxel size, and channels, not the benefit of optimizing L-mAP per se. The authors should evaluate L-HPO on a held-out split, compare against a CenterPoint variant tuned for mAP with the same budget, and report the selected hyperparameters.
minor comments (6)
  1. [Abstract and Sec. 4.3] The phrase 'state-of-the-art performance model' is vague; the paper shows one tuned model but does not compare against other HPO or NAS methods for real-time 3D detection.
  2. [Sec. 4.2] The text states that 'mAP remains consistently at 100.0' for the latency and yaw experiments, but Table 3(a) and (b) do not include mAP columns; please report those numbers explicitly.
  3. [Sec. 3.1, Eq. (3)] The subtraction of v_ego_t * Δt in Eq. (3) requires a clear statement of the coordinate frame: if x_pred_t is in the ego-centric frame, the correction is appropriate, but if it is in a global frame, it is not; please specify the frame.
  4. [Sec. 4.4, Table 6] The caption says 'PN is the number of merged point clouds' but the context implies the number of LiDAR points; please clarify the notation.
  5. [Related Work] Given the similarity to Pl-metrics [18], the paper should provide a direct comparison or clear differentiation between P-AP and the planning cost function used in [18].
  6. [Appendix, Sec. 6.2] Equation (6) appears to have a typo or inconsistency in the expansion; please verify the algebra for the position estimation error.

Circularity Check

1 steps flagged · score 6.0 of 10

L-HPO's state-of-the-art L-mAP is a same-split fit: hyperparameters are optimized to maximize L-mAP on the nuScenes validation set and then reported as surpassing other models on that same validation set, making the main performance claim partly a restatement of the optimization objective.

  1. fitted input called prediction [Sec. 3.3 (Latency-aware Hyperparameter Optimization) and Sec. 4.3 (Benchmark with Our Metrics), Table 4]
    "In this paper, we use the CenterPoint [51] model as base model to optimize hyperparameters such as the number of layers, the number of channels in each layer, the number of multi-frame to merge point clouds, and the voxel size, aiming to find the model for the highest L-mAP. ... From Sec. 4.3 to Sec. 4.5, we evaluate 3D object detection based on our metrics using the validation dataset of the nuScenes dataset. ... L-HPO outperforms other models in L-mAP and demonstrates high performance for real-time 3D object detection applications. Specifically, L-HPO increase L-mAP by +8.4 from CenterPoint."

    L-HPO's hyperparameters are explicitly selected to maximize L-mAP, and the evaluation in Table 4 is performed on the same nuScenes validation split used for that selection. No held-out or test split is reported for L-HPO. Consequently, the headline result — L-HPO's +8.4 L-mAP advantage over CenterPoint and its 'state-of-the-art' characterization — is the direct output of optimizing the reported objective on the reported dataset, rather than an independent prediction of performance. The claim 'L-HPO outperforms other models in L-mAP' restates the optimization target as an experimental finding.

full rationale

The paper's core metric proposals (L-AP and P-AP) are not circular in the narrow sense: L-AP is defined by shifting detections with velocity times latency, and the nuPlan closed-loop correlation in Sec. 4.2 is an external, empirically measured relationship, even if its synthetic-error methodology limits external validity. That limitation is a correctness risk, not a circularity. There are no load-bearing self-citations: the references to CenterPoint, TransFusion, BEVFusion, etc. are standard prior work, and no uniqueness or ansatz is imported from the authors' own earlier papers. The main circularity is the L-HPO benchmark: the same nuScenes validation split is used both for hyperparameter optimization (maximizing L-mAP) and for the 'state-of-the-art' L-mAP comparison. Because the reported metric is identical to the objective being optimized, the central performance claim partially reduces to a fit. If L-HPO had been evaluated on a held-out test split or on nuPlan, the paper would be substantially more independent; as written, the SOTA claim is a same-split fitted result, warranting a score of 6.

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

The central claims rest on the constant-velocity and ego-motion assumptions inside L-AP, on the hand-set planning margin and occlusion filter inside P-AP, and on the synthetic-error validation methodology that stands in for real detector errors. The L-HPO result additionally depends on undisclosed hyperparameters fitted and evaluated on the same validation split.

free parameters (3)
  • Planning margin dm = 0.5 m
    Hand-set in Sec. 3.2 for P-AP matching; no sensitivity analysis is provided, and all P-AP scores depend on this value.
  • P-mAP distance thresholds = (0.5, 1.0, 1.5, 2.0) m
    Chosen thresholds in Table 1 and Sec. 3.2; changing them changes P-mAP scores and model rankings.
  • L-HPO hyperparameter configuration = Not disclosed
    Number of layers, channels, multi-frame count, and voxel size were searched to maximize L-mAP on the nuScenes validation split, and the final model was evaluated on the same split.
assumptions (5)
  • domain assumption Objects move with roughly constant velocity over the inference latency interval (Eqs. 2 and 3).
    L-AP shifts predicted and ground-truth positions linearly in time using velocities, ignoring acceleration and jerk in the main metric.
  • domain assumption The ego-motion compensation in Eq. (3) is the correct transformation for comparing detections at inference completion.
    The paper subtracts ego velocity from predicted positions but not from ground-truth positions, and it does not state the coordinate frame in which the comparison is made.
  • ad hoc to paper The planning margin dm=0.5 m is a valid fixed proxy for safe planning behavior.
    P-AP treats any farther-than-truth nearest-surface error larger than dm as an unsafe non-match; this is a design choice without empirical calibration.
  • ad hoc to paper Occluded objects can be filtered out because they do not significantly affect motion planning.
    Sec. 3.2 excludes occluded objects from P-AP, but the paper does not define an occlusion criterion or justify why occluded objects never matter for planning.
  • ad hoc to paper Injecting latency, yaw, and position errors into ground-truth detections approximates the error distribution of real 3D detectors.
    Sec. 4.2 uses ground truth as predictions to demonstrate that L-mAP and P-mAP correlate with nuPlan score; real detector errors may be spatially correlated and differently distributed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Rethink 3D Object Detection from Physical World." pith.science (2026). https://pith.science/paper/3D63NNIR

@misc{pith2026250700190,
  author       = {Pith},
  title        = {Pith review of: Rethink 3D Object Detection from Physical World},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3D63NNIR}},
  note         = {Machine review of arXiv:2507.00190}
}
read the original abstract

High-accuracy and low-latency 3D object detection is essential for autonomous driving systems. While previous studies on 3D object detection often evaluate performance based on mean average precision (mAP) and latency, they typically fail to address the trade-off between speed and accuracy, such as 60.0 mAP at 100 ms vs 61.0 mAP at 500 ms. A quantitative assessment of the trade-offs between different hardware devices and accelerators remains unexplored, despite being critical for real-time applications. Furthermore, they overlook the impact on collision avoidance in motion planning, for example, 60.0 mAP leading to safer motion planning or 61.0 mAP leading to high-risk motion planning. In this paper, we introduce latency-aware AP (L-AP) and planning-aware AP (P-AP) as new metrics, which consider the physical world such as the concept of time and physical constraints, offering a more comprehensive evaluation for real-time 3D object detection. We demonstrate the effectiveness of our metrics for the entire autonomous driving system using nuPlan dataset, and evaluate 3D object detection models accounting for hardware differences and accelerators. We also develop a state-of-the-art performance model for real-time 3D object detection through latency-aware hyperparameter optimization (L-HPO) using our metrics. Additionally, we quantitatively demonstrate that the assumption "the more point clouds, the better the recognition performance" is incorrect for real-time applications and optimize both hardware and model selection using our metrics.

Figures

Figures reproduced from arXiv: 2507.00190 by the authors.

Figure 1
Figure 1. The cases where a high score on existing mAP does not correspond to practical performance. Grey objects represent ground truths [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The metric of L-AP. L-AP measures the accuracy of [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (a) Mis-detection towards unsafe and safe. If a 3D object detection model misidentifies objects as being farther away than their [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: The case to evaluate of corner distance. Grey objects [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: The case to evaluate for yaw flipping. (a) The velocity [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: The estimation error of x GT t+∆t for inference time ∆t = 0.2 s. estimation error. Conversely, given the inference time δt and acceptable error in a system, the required annotation interval can be determined as a specification requirement. Next, we analyze the estimati…
Figure 7
Figure 7. Figure 7: The estimation error of x GT t+∆t for inference time ∆t = 0.5 s. (a) δt = 1.0 (1 Hz annotation). (b) δt = 0.5 (2 Hz annotation). (c) δt = 0.2 (5 Hz annotation). (d) δt = 0.1 (10 Hz annotation) [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: The estimation error of v GT t+∆t for inference time ∆t = 0.2 s. 3 [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: The estimation error of v GT t+∆t for inference time ∆t = 0.5 s. Annotation Ground Truth t = T t = T + dt V gt V an = (x(t+dt) - xt) / dt Point clouds [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: The error of annotation. annotation position error is 0.5 m and the annotation inter￾val is 0.5 s (2 Hz annotation), the resulting velocity estima￾tion error can reach 1.0 m/s. Therefore, it is crucial to take such annotation-induced errors into account when evaluat￾i…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 46 canonical work pages

  1. [1]

    https://github.com/NVIDIA- AI- IOT/Lidar_AI_Solution

    Nvidia-ai-iot. https://github.com/NVIDIA- AI- IOT/Lidar_AI_Solution. 6

    work page

  2. [2]

    https : / / github

    Tensorrt. https : / / github . com / NVIDIA / TensorRT. 5

    work page

  3. [3]

    Yolo3d: End-to-end real-time 3d oriented object bounding box detection from li- dar point cloud

    Waleed Ali, Sherif Abdelkarim, Mohamed Zahran, Mah- moud Zidan, and Ahmad El Sallab. Yolo3d: End-to-end real-time 3d oriented object bounding box detection from li- dar point cloud. In ECCV Workshops, 2018. 3

    work page 2018

  4. [4]

    Transfusion: Robust lidar-camera fusion for 3d object detection with transform- ers

    Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transform- ers. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1080–1089, 2022. 2, 6, 7

    work page 2022

  5. [5]

    Multicorrupt: A multi-modal robustness dataset and benchmark of lidar-camera fusion for 3d object detec- tion

    Till Beemelmanns, Quan Zhang, Christian Geller, and Lutz Eckstein. Multicorrupt: A multi-modal robustness dataset and benchmark of lidar-camera fusion for 3d object detec- tion. In 2024 IEEE Intelligent Vehicles Symposium (IV) , pages 3255–3261, 2024. 3

    work page 2024

  6. [6]

    Omni3D: A large benchmark and model for 3D object detection in the wild

    Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3D: A large benchmark and model for 3D object detection in the wild. In CVPR, Vancouver, Canada, 2023. IEEE. 3

    work page 2023

  7. [7]

    Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom

    Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom. nuscenes: A multi- modal dataset for autonomous driving. In CVPR, 2020. 1, 3, 5

    work page 2020

  8. [8]

    Evaluating ob- ject (mis)detection from a safety and reliability perspective: Discussion and measures

    Andrea Ceccarelli and Leonardo Montecchi. Evaluating ob- ject (mis)detection from a safety and reliability perspective: Discussion and measures. IEEE Access, 11:44952–44963,

    work page

Show all 54 references
  1. [9]

    Wellington

    Siheng Chen, Baoan Liu, Chen Feng, Carlos Vallespi- Gonzalez, and Carl K. Wellington. 3d point cloud processing and learning for autonomous driving: Impacting map cre- ation, localization, and perception. IEEE Signal Processing Magazine, 38:68–86, 2020. 2

    2020

  2. [10]

    Multi-view 3d object detection network for autonomous driving

    Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. 2017 IEEE Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 6526–6534, 2016. 2

    2017

  3. [11]

    Futr3d: A unified sensor fusion framework for 3d detection

    Xuanyao Chen, Tianyuan Zhang, Yue Wang, Yilun Wang, and Hang Zhao. Futr3d: A unified sensor fusion framework for 3d detection. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 172–181, 2022. 3

    2023

  4. [12]

    Parting with misconceptions about learning- based vehicle motion planning

    Daniel Dauner, Marcel Hallgarten, Andreas Geiger, and Kashyap Chitta. Parting with misconceptions about learning- based vehicle motion planning. In Conference on Robot Learning (CoRL), 2023. 5

    2023

  5. [13]

    Deepfusion: A ro- bust and modular 3d object detector for lidars, cameras and radars

    Florian Drews, Di Feng, Florian Faion, Lars Rosenbaum, Michael Ulrich, and Claudius Gl ¨aser. Deepfusion: A ro- bust and modular 3d object detector for lidars, cameras and radars. 2022 IEEE/RSJ International Conference on Intelli- gent Robots and Systems (IROS), pages 560–567, 2022. 2

    2022

  6. [14]

    Bevdet: High-performance multi-camera 3d object detection in bird-eye-view

    Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. ArXiv, abs/2112.11790, 2021. 3

    2021 arXiv

  7. [15]

    Detecting as labeling: Rethinking lidar-camera fusion in 3d object detection

    Junjie Huang, Yun Ye, Zhujin Liang, Yi Shan, and Dalong Du. Detecting as labeling: Rethinking lidar-camera fusion in 3d object detection. In Computer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29 – October 4, 2024, Proceedings, Part XXII, page 439–45...

    2024

  8. [16]

    Multi-modal sensor fusion for auto driving perception: A survey

    Keli Huang, Botian Shi, Xiang Li, Xin Li, Siyuan Huang, and Yikang Li. Multi-modal sensor fusion for auto driving perception: A survey. ArXiv, abs/2202.02703, 2022. 2

    2022 arXiv

  9. [17]

    Let-3d-ap: Longitudi- nal error tolerant 3d average precision for camera-only 3d de- tection

    Wei-Chih Hung, Vincent Casser, Henrik Kretzschmar, Jyh- Jing Hwang, and Dragomir Anguelov. Let-3d-ap: Longitudi- nal error tolerant 3d average precision for camera-only 3d de- tection. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 8272–8279, 2024. 3

    2024

  10. [18]

    Ivanovic and Marco Pavone

    B. Ivanovic and Marco Pavone. Injecting planning- awareness into prediction and detection evaluation. 2022 IEEE Intelligent Vehicles Symposium (IV) , pages 821–828,

    2022

  11. [19]

    Far3d: Expanding the horizon for surround-view 3d object detec- tion

    Xiaohui Jiang, Shuailin Li, Yingfei Liu, Shihao Wang, Fan Jia, Tiancai Wang, Lijin Han, and Xiangyu Zhang. Far3d: Expanding the horizon for surround-view 3d object detec- tion. In Thirty-Eighth AAAI Conference on Artificial Intel- ligence, AAAI 2024 , pages 2561–2569. AAAI Pre...

    2024

  12. [20]

    Towards learning- based planning: The nuplan benchmark for real-world au- tonomous driving

    Napat Karnchanachari, Dimitris Geromichalos, Kok Seang Tan, Nanxiang Li, Christopher Eriksen, Shakiba Yaghoubi, Noushin Mehdipour, Gianmarco Bernasconi, Whye Kit Fong, Yiluan Guo, and Holger Caesar. Towards learning- based planning: The nuplan benchmark for real-world au- tono...

    2024

  13. [21]

    Lang, Sourabh V ora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom

    Alex H. Lang, Sourabh V ora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12689–12697, 2018. 2, 6, 7

    2019

  14. [22]

    Qiao, Dahua Lin, Siqian Liu, Junchi Yan, Jianping Shi, and Ping Luo

    Hongyang Li, Chonghao Sima, Jifeng Dai, Wenhai Wang, Lewei Lu, Huijie Wang, Enze Xie, Zhiqi Li, Hanming Deng, Haonan Tian, Xizhou Zhu, Li Chen, Tianyu Li, Yulu Gao, Xiangwei Geng, Jianqiang Zeng, Yang Li, Jiazhi Yang, Xi- aosong Jia, Bo Yu, Y . Qiao, Dahua Lin, Siqian Liu, Jun...

    2022

  15. [23]

    Bevdepth: Acquisition of reliable depth for multi-view 3d object detection

    Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zen- gran Wang, Yukang Shi, Jian-Yuan Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. ArXiv, abs/2206.10092, 2022. 3

    2022 arXiv

  16. [24]

    Le, Alan Yuille, and Mingxing Tan

    Yingwei Li, Adams Wei Yu, Tianjian Meng, Ben Caine, Ji- quan Ngiam, Daiyi Peng, Junyang Shen, Yifeng Lu, Denny Zhou, Quoc V . Le, Alan Yuille, and Mingxing Tan. Deep- fusion: Lidar-camera deep fusion for multi-modal 3d object detection. In 2022 IEEE/CVF Conference on Computer ...

    2022

  17. [25]

    Fully sparse fusion for 3d object detection

    Yingyan Li, Lue Fan, Yang Liu, Zehao Huang, Yuntao Chen, Naiyan Wang, and Zhaoxiang Zhang. Fully sparse fusion for 3d object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2

    2024

  18. [26]

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

    Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chong- hao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers. pages 1–18, 2022. 3, 6, 7

    2022

  19. [27]

    Bevfusion: A simple and robust lidar-camera fusion framework

    Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. ArXiv, abs/2205.13790, 2022. 2

    2022 arXiv

  20. [28]

    Qi, Yin Zhou, Mingxing Tan, and Dragomir Anguelov

    Chenxi Liu, Zhaoqi Leng, Pei Sun, Shuyang Cheng, Charles R. Qi, Yin Zhou, Mingxing Tan, and Dragomir Anguelov. Lidarnas: Unifying and searching neural archi- tectures for 3d point clouds. In Computer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27...

    2022

  21. [29]

    Bevfusion: Multi- task multi-sensor fusion with unified bird’s-eye view repre- sentation

    Zhijian Liu, Haotian Tang, Alexander Amini, Xingyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multi- task multi-sensor fusion with unified bird’s-eye view repre- sentation. In IEEE International Conference on Robotics and Automation (ICRA), 2023. 2, 6, 7

    2023

  22. [30]

    3d object detection for autonomous driving: A comprehensive survey, 2023

    Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hong- sheng Li. 3d object detection for autonomous driving: A comprehensive survey, 2023. 2

    2023

  23. [31]

    Mtbf model for avs - from per- ception errors to vehicle-level failures.2022 IEEE Intelligent Vehicles Symposium (IV), pages 1591–1598, 2022

    Fabian Oboril, Cornelius B ¨urkle, Alon Sussmann, Simcha Bitton, and Simone Fabris. Mtbf model for avs - from per- ception errors to vehicle-level failures.2022 IEEE Intelligent Vehicles Symposium (IV), pages 1591–1598, 2022. 3

    2022

  24. [32]

    Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unproject- ing to 3d

    Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unproject- ing to 3d. In Computer Vision – ECCV 2020: 16th Euro- pean Conference, Glasgow, UK, August 23–28, 2020, Pro- ceedings, Part XIV, page 194–210, Berlin, Heidel...

    2020

  25. [33]

    Qi, Hao Su, Kaichun Mo, and Leonidas J

    C. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Point- net: Deep learning on point sets for 3d classification and seg- mentation. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 77–85, 2016. 2

    2017

  26. [34]

    Qi, Li Yi, Hao Su, and Leonidas J

    Charles R. Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Point- net++: deep hierarchical feature learning on point sets in a metric space. In Proceedings of the 31st International Conference on Neural Information Processing Systems, page 5105–5114, Red Hook, NY , USA, 2017. Curra...

    2017

  27. [35]

    Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J

    Charles R. Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J. Guibas. Frustum pointnets for 3d object de- tection from rgb-d data. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 918–927,

    2018

  28. [36]

    Pointr- cnn: 3d object proposal generation and detection from point cloud

    Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr- cnn: 3d object proposal generation and detection from point cloud. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–779, 2018. 2

    2019

  29. [37]

    Pv-rcnn: Point-voxel feature set abstraction for 3d object detection

    Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jian- ping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. 2020 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 10526–10535, 2019. 2

    2020

  30. [38]

    Scalability in perception for autonomous driving: Waymo open dataset

    Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Et- tinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhan...

    2020

  31. [39]

    Fcos: Fully convolutional one-stage object detection

    Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9626–9635, 2019. 3

    2019

  32. [40]

    Lang, Bassam Helou, and Oscar Bei- jbom

    Sourabh V ora, Alex H. Lang, Bassam Helou, and Oscar Bei- jbom. Pointpainting: Sequential fusion for 3d object detec- tion. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4603–4611, 2019. 2

    2020

  33. [41]

    Pointaugmenting: Cross-modal augmentation for 3d object detection

    Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 11794– 11803, 2021. 2

    2021

  34. [42]

    Towards stable 3d object detection

    Jiabao Wang, Qiang Meng, Guochao Liu, Liujiang Yan, Ke Wang, Ming-Ming Cheng, and Qibin Hou. Towards stable 3d object detection. In Computer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29–October 4, 2024, Proceedings, Part L, page 197–213, Berlin, Hei...

    2024

  35. [43]

    Exploring object-centric temporal modeling for efficient multi-view 3d object detection

    Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xi- angyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3598–3608, 2023. 3

    2023

  36. [44]

    Xiaofang Wang and Kris M. Kitani. Cost-aware evaluation and model scaling for lidar-based 3d object detection. 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9260–9267, 2022. 3

    2023

  37. [45]

    Yue Wang, Vitor Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, , and Justin M. Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In The Conference on Robot Learning (CoRL), 2021. 3

    2021

  38. [46]

    Second: Sparsely embed- ded convolutional detection

    Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embed- ded convolutional detection. Sensors (Basel, Switzerland) , 18, 2018. 2

    2018

  39. [47]

    Pixor: Real- time 3d object detection from point clouds

    Binh Yang, Wenjie Luo, and Raquel Urtasun. Pixor: Real- time 3d object detection from point clouds. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 7652–7660, 2018. 2 10

    2018

  40. [48]

    BEVFormer v2: Adapt- ing Modern Image Backbones to Bird’s-Eye-View Recogni- tion via Perspective Supervision

    Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, Jie Zhou, and Jifeng Dai. BEVFormer v2: Adapt- ing Modern Image Backbones to Bird’s-Eye-View Recogni- tion via Perspective Supervision . pages 17830–17839, 2023. 3

    2023

  41. [49]

    3dssd: Point-based 3d single stage object detector

    Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11037–11045, 2020. 2

    2020

  42. [50]

    Zetong Yang, Zhiding Yu, Chris Choy, Renhao Wang, Anima Anandkumar, and Jose M. Alvarez. Improving distant 3d ob- ject detection using 2d box supervision. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14853–14863, 2024. 3

    2024

  43. [51]

    Center- based 3d object detection and tracking

    Tianwei Yin, Xingyi Zhou, and Philipp Kr ¨ahenb¨uhl. Center- based 3d object detection and tracking. CVPR, 2021. 2, 5, 6, 7

    2021

  44. [52]

    Beverse: Uni- fied perception and prediction in birds-eye-view for vision- centric autonomous driving

    Yunpeng Zhang, Zheng Hua Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Uni- fied perception and prediction in birds-eye-view for vision- centric autonomous driving. ArXiv, abs/2205.09743, 2022. 3

    2022 arXiv

  45. [53]

    V oxelnet: End-to-end learning for point cloud based 3d object detection

    Yin Zhou and Oncel Tuzel. V oxelnet: End-to-end learning for point cloud based 3d object detection. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 4490–4499, 2017. 2 11 Rethink 3D Object Detection from Physical World Supplementary Material

    2018

  46. [54]

    L-AP for Yaw Flipping 10m/s GT(a) 13m/s (b) -7m/s 1m/s GT(c) 1.3m/s (d) -0.7m/s Figure 5

    Appendix 6.1. L-AP for Yaw Flipping 10m/s GT(a) 13m/s (b) -7m/s 1m/s GT(c) 1.3m/s (d) -0.7m/s Figure 5. The case to evaluate for yaw flipping. (a) The velocity estimation is sufficiently accurate. (b) The direction is estimated in the opposite orientation. (c) The object is al...

Pith tools

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