Pith. sign in

REVIEW 5 major objections 5 minor 52 references

Towards Accurate and Efficient 3D Object Detection for Autonomous Driving: A Mixture of Experts Computing System on Edge

T0 review · 5 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 EMC2 claims that routing LiDAR-camera scenes among three specialized experts lets an autonomous-driving detector beat 15 baselines on accuracy and latency simultaneously.

desk verdict A plausible MoE system for edge 3D detection whose headline numbers are undermined by single-baseline speedup statistics, validation-set tuning, and an ablation narrative that contradicts its own table. read the letter →

arxiv 2507.04123 v2 pith:6EQUANEU submitted 2025-07-05 cs.CV cs.AI

Linshen Liu , Boyan Su , Junyue Jiang , Guanlin Wu , Cong Guo , Ceyu Xu , Hao Frank Yang This is my paper
classification cs.CVcs.AI
keywords 3Dobjectdetectionmixtureofexpertsscenario-awareroutingmultimodalfusionLiDARandcameraedgecomputingJetsonAGXOrinautonomousdriving
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 tries to establish that autonomous driving perception can escape the usual accuracy-versus-latency trade-off by replacing one uniform detector with a mixture of scenario-specialized experts. It claims that EMC2, an edge-oriented multimodal system, improves 3D detection accuracy on KITTI by an average of 3.58% while reducing inference latency by 159.06% compared with 15 baseline detectors on Jetson platforms, and shows similar gains on nuScenes. The stakes are practical: an autonomous vehicle must both recognize distant, occluded, or hard objects reliably and react within about 100 milliseconds, and edge hardware makes meeting both demands hard. The paper's solution is to route each scene to a fast 2D expert for close, clear cases, a 3D LiDAR expert for mixed cases, and a full LiDAR-plus-image expert for distant uncertain cases, then optimize the resulting graph for edge execution.

What carries the argument

The load-bearing object is the Scenario-Adaptive Dispatcher (SAD), a router that reads proposal regions from the Adaptive Multimodal Data Bridge and classifies each frame by object distance and confidence. If all proposals lie within a distance threshold D and have confidence at least C, the frame goes to the Latency-Prioritized Expert, a 2D CNN operating on bird's-eye-view projections; if some objects are distant or unclear, it goes to the Versatile Efficiency Expert, a sparse 3D CNN; if any object is both distant and uncertain, it goes to the Accuracy-Prioritized Expert, which fuses image features into LiDAR voxel features via multiscale pooling. The thresholds are set empirically at 23.5 meters for KITTI and 35 meters for nuScenes. The dispatcher is what makes the system's average cost a mix of cheap and expensive experts, and the hierarchical triple back-propagation scheme is what keeps the mixture trainable when proposals are unreliable early in training.

What would settle it

Take a held-out split of KITTI or nuScenes, run every scene through all three experts, and compare the dispatcher's chosen expert with the oracle best expert for that scene; if the router is wrong on a substantial share of scenes, the reported accuracy gain would shrink toward the weaker expert's performance, and the speedup could be reproduced on easy scenes by using the cheap expert alone, which is the baseline the paper does not provide.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that EMC2 is the first system to make MoE-based multimodal 3D detection both more accurate and faster than one-size-fits-all baselines on edge devices. On KITTI, it reports a 159.06% inference speedup on the Jetson AGX Orin and accuracy that is higher than all 15 compared methods in most difficulty levels, with the largest margins on hard-level pedestrians and cyclists; on nuScenes it reports mAP 0.7241 and NDS 0.7316 at 229.3 ms, ahead of BEVFusion and FocalFormer. The authors do not attribute this to one novel network, but to the combination of a scenario-aware dispatcher, three complementary experts, hierarchical training with balanced sampling and adaptive learning rates, and system-level optimizations including custom sparse convolution, multiscale pooling, graph fusion, and staged thread management.

Load-bearing premise

The load-bearing assumption is that the router's fixed distance and confidence cutoffs, chosen by looking at validation data, keep selecting the right expert on scenes the system has never seen; the paper never directly measures how often the router picks the wrong expert.

Editorial extensions

If this is right

  • The reported results imply that real-time multimodal 3D detection is feasible on a Jetson AGX Orin at 372.5 ms per frame on KITTI, while LiDAR-only baselines on the same platform need 965–1787 ms.
  • Average latency becomes scene-dependent: most close-range frames use the cheap 2D expert, and the expensive image-fusing expert is reserved for distant or low-confidence objects, so the cost scales with difficulty rather than with model size.
  • The ablation claims that hierarchical training and multiscale pooling are jointly required for the gains: dropping multiscale pooling raises latency from 526 ms to 1320 ms, and dropping hierarchical training lowers accuracy on all three KITTI classes.
  • Because the system is exported to ONNX, the same dispatch-and-optimize strategy can be ported to other edge platforms with different compilers, not only the Jetson hardware tested here.

Reading between the lines

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

  • The router's distance and confidence thresholds are chosen by inspecting validation data, and the paper does not report a direct measure of routing accuracy on unseen scenes, so the least-tested component is whether the dispatcher generalizes across cities, sensor setups, and object classes.
  • A learned or online-adaptive router would be the natural next step; if routing were trained end-to-end against detection loss, the fixed 23.5 m and 35 m thresholds could become contextual rather than dataset-specific.
  • The reported average latency depends on the mix of easy and hard scenes in the benchmark; on a distribution with more distant and occluded objects, EMC2's average cost would shift toward the 526 ms multimodal expert, narrowing the speedup.
  • One testable consequence of the paper's design is that the accuracy gain should be explainable by oracle routing: running the oracle best expert on every scene should match or exceed EMC2, and the distance between EMC2 and the oracle bounds how much routing is leaving on the table.
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 / 5 minor

Summary. The paper proposes EMC2, an edge-oriented mixture-of-experts (MoE) computing system for multimodal 3D object detection in autonomous driving. The system combines an Adaptive Multimodal Data Bridge (AMDB), a Scenario-Adaptive Dispatcher (SAD) that routes scenes to one of three experts (Latency-Prioritized, Versatile Efficiency, Accuracy-Prioritized) based on object distance and confidence, and a set of system-level optimizations for Jetson-class hardware. The authors report state-of-the-art accuracy and latency on KITTI and nuScenes, claiming a 159.06% inference speedup over 15 baselines on Jetson platforms. The manuscript also describes a hierarchical training strategy with three back-propagation paths, a balanced sampling scheme, and multiscale pooling.

Significance. The idea of scenario-aware MoE for edge 3D detection is timely and practically motivated, and the paper represents a genuine engineering effort: it releases code, targets a specific edge platform (Jetson AGX Orin), and combines algorithm-level and system-level optimizations (custom sparse convolution, multiscale pooling, memory management, graph fusion). If the reported results were valid, EMC2 would be a significant demonstration of real-time multimodal detection on edge devices. However, the manuscript's central quantitative claims are not supported by the data as presented, and the validation methodology contains a fundamental circularity. The strengths of the code release and hardware co-design cannot compensate for these issues, so the significance of the contribution is not currently established.

major comments (5)
  1. [Abstract, Sec. 4.2, Table 1] The abstract's claim of a 159.06% inference speedup 'compared to 15 baseline methods on Jetson platforms' is not supported. In Table 1, only five of the fifteen baselines have reported Jetson latency (PV-RCNN 1787 ms, PartA2-Free 1064 ms, PointPillar 972 ms, Second 1322 ms, and multimodal Voxel-RCNN 965 ms); the remaining ten are marked N/A. The 159.06% figure is exactly (965/372.5 - 1), i.e., a comparison against the single slowest measured baseline. Averaging the speedup over the five available baselines gives roughly a 2.28x speedup (about 128%), not 2.59x. The headline speedup therefore misrepresents the actual comparison, and the 'average accuracy improvement of 3.58%' is similarly undefined given the many N/A entries. This is a load-bearing error: the paper's central claim of simultaneous accuracy and efficiency rests on this number.
  2. [Sec. 4.1 and Sec. 4.3] The routing thresholds D (23.5 m on KITTI, 35 m on nuScenes) and the confidence threshold C are chosen by inspecting validation-set proposal-confidence distributions, and the routing is then validated on the same validation set. Figure 7 labels validation-set distance/confidence bins with the empirically best-performing expert and treats this as evidence that the router works. This is circular: the routing rule is fit to the validation set and then read back from it. The paper never reports routing accuracy on held-out scenes, and the confidence threshold C is not even reported. Because the accuracy-speedup tradeoff depends critically on correct expert selection, the claimed generalization to unseen scenes is not established.
  3. [Sec. 4.4, Table 3] The text states that 'LPE, VEE, and APE achieve comparable accuracy in LPE scenes,' but Table 3 shows LPE at 66.32/89.04/88.54 for Ped/Car/Cyclist versus VEE at 74.49/92.23/91.69 and APE at 74.34/93.21/92.41. For pedestrians, the gap is roughly 8 AP points, which is not 'comparable.' This contradiction undermines the premise that the Latency-Prioritized Expert can handle close-and-clear scenes without meaningful accuracy loss, which is the basis for the claimed efficiency gain.
  4. [Sec. 3.4] The training description is internally inconsistent. The text describes hierarchical training with three back-propagation pathways (from the LiDAR branch, the image branch, and expert predictions), but then states that during joint training 'gradients are propagated only from APE, which receives full multimodal input.' As written, this means LPE and VEE do not receive gradients from their own outputs during joint training, leaving unclear how they are actually trained end-to-end. This is a load-bearing detail for the accuracy results, and the description is not reproducible as it stands.
  5. [Sec. 4.4, Table 3 (HT/MP ablation)] The ablation paragraph misreports the table. It says 'For the car class, MP alone reduces accuracy from 80.83% to 75.98%,' but Table 3 shows these numbers are the cyclist column (HT✗ MP✗: 80.83; HT✗ MP✓: 75.98), while the car column goes from 89.10 to 88.68. It then says 'HT alone improves it from 66.34% to 70.55%,' which is the pedestrian column, not the car column. The text mislabels the classes, making the ablation discussion unreliable as a description of the results.
minor comments (5)
  1. [Sec. 4.2] The sentence 'It runs at 372.5 ms on Jetson, 2.5×–2.5× faster than methods requiring 965–1,787 ms' contains a nonsensical range; the speedup relative to 965 ms is about 2.59x and relative to 1787 ms is about 4.80x, so the range should be stated accordingly.
  2. [General / Tables 1-3] No error bars, standard deviations, or multiple-run statistics are reported for any AP or latency number, so it is impossible to judge the statistical significance of the claimed improvements.
  3. [Fig. 2 and Sec. 3] The 'Emergency Expert API' appears in the system diagram but is never described or evaluated anywhere in the text; either it should be removed or its role should be specified.
  4. [Table 2 and Sec. 4.2] The nuScenes comparison includes only two baselines (FocalFormer and BEVFusion), and neither has a reported Jetson latency, so the abstract's suggestion of 'similar performance gains' on nuScenes is not established with the same breadth as the KITTI claim.
  5. [Sec. 4.1] The paper evaluates on the KITTI validation set rather than the official test set; this limits comparability with published SOTA numbers, most of which are test-set results, and should be stated prominently.

Circularity Check

1 steps flagged · score 6.0 of 10

Dispatcher validation reduces to a self-fit: the validation grid is labeled by the best expert per bin, so the routing rule's agreement with it is a restatement of the grid construction; the end-to-end benchmark tables remain independent evidence.

  1. fitted input called prediction [Section 4.1 and Section 4.3 / Fig. 7]
    "A two-sample Kolmogorov–Smirnov test reveals a statistically significant difference between objects within and beyond a distance threshold D, which is empirically set to 23.5 meters for KITTI and 35 meters for nuScenes. ... This strategy is validated via expert activations on the nuScenes validation set (Fig. 7): LPE performs best in close-range, high-confidence scenes; APE in distant, low-confidence cases; and VEE in all others, supporting the effectiveness of empirical routing. ... Each cell is labeled with the name of the expert that performs best for scenes falling into that bin."

    The routing thresholds are empirical: D is 'empirically set' from the data, and the confidence threshold C is never reported. The purported validation then takes the same validation set and colors each distance/confidence bin with the expert that 'performs best' in that bin. Because the validation grid is constructed as the best-expert oracle, the observed pattern (LPE for close/high-confidence, APE for distant/low-confidence, VEE otherwise) is a restatement of the grid's definition, not an independent test of whether EMC2's fixed thresholds select that expert. The paper presents this as 'supporting the effectiveness of empirical routing' without ever measuring routing accuracy directly on held-out scenes, so the validation reduces to a self-fit of the routing rule.

full rationale

The central end-to-end accuracy and latency comparisons in Tables 1 and 2 are measured against external baselines (KITTI and nuScenes) and are not themselves derived from the routing validation; those numbers remain independent evidence. The specific circularity is localized to the dispatcher validation in Section 4.3 / Fig. 7: the validation grid is defined by labeling each bin with the best-performing expert, and the paper then reads the resulting pattern as confirmation of the empirically set routing thresholds. This is a construction-level tautology rather than a predictive test. The confidence threshold C is also unreported, making it impossible to check whether the fixed routing rule matches the oracle grid without post hoc adjustment. No load-bearing self-citations, imported uniqueness theorems, or ansatz-by-citation issues appear, and the claimed 159.06% speedup aggregation problem is a claim-support/reporting concern rather than a circularity. On balance, one central supporting 'validation' reduces by construction, but the measured benchmark results retain independent content, giving partial circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central accuracy and latency claim rests on manually chosen thresholds, a simple routing assumption contradicted in part by the ablation, and a training scheme whose gradient flow is underspecified. These are the costs a reader pays beyond standard supervised detection.

free parameters (4)
  • Distance threshold D = 23.5 m (KITTI), 35 m (nuScenes)
    Set empirically after a KS test on validation proposal confidences (Sec 4.1). It directly decides which expert handles a scene.
  • Confidence threshold C = not reported
    Used by the Scenario-Adaptive Dispatcher to separate clear from unclear proposals (Sec 3.2); the paper says thresholds are fixed and dataset-adapted but never gives C.
  • Multiscale pooling size = not specified
    Described as user-configurable (Sec 3.5); affects memory usage, latency, and fusion quality, but no value or sensitivity analysis is given.
  • Base learning rate alpha_0 = not reported
    Appears in the adaptive optimizer formula (Eq 3); its value and schedule are not given, so exact reproduction is impossible.
assumptions (4)
  • domain assumption Close, high-confidence objects can be detected as accurately by a 2D BEV CNN as by 3D or multimodal experts.
    This is the basis for routing to LPE (Sec 3.2). Table 3 actually shows LPE with pedestrian AP 66.32 vs 74.49 for VEE and 74.34 for APE on LPE target scenes, so the assumption is not supported by the paper's own numbers.
  • domain assumption Distance and AMDB proposal confidence are sufficient routing features for unseen scenes.
    The SAD dispatcher uses only D and C (Sec 3.2); no routing accuracy or misrouting cost is measured. The expert-activation map (Fig 7) is built from validation data.
  • domain assumption Validation splits used to set thresholds are representative of deployment conditions.
    D is chosen from KITTI and nuScenes validation proposal-confidence distributions (Sec 4.1); there is no test on a different city, sensor, or weather condition.
  • ad hoc to paper Backpropagating only from APE during joint training is sufficient to train LPE and VEE.
    Sec 4.1 states AMDB is pre-trained and then gradients propagate only from APE; how LPE and VEE heads are trained is not specified, yet the central accuracy claim depends on all three experts working.
invented entities (1)
  • Emergency Expert API
    purpose: A module for rapid response to hazardous or unseen situations, shown in Fig 2 and described in the Sec 3.3 caption.
    It is never implemented, evaluated, or precisely defined; it is a placeholder interface rather than a functioning component.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Accurate and Efficient 3D Object Detection for Autonomous Driving: A Mixture of Experts Computing System on Edge." pith.science (2026). https://pith.science/paper/6EQUANEU

@misc{pith2026250704123,
  author       = {Pith},
  title        = {Pith review of: Towards Accurate and Efficient 3D Object Detection for Autonomous Driving: A Mixture of Experts Computing System on Edge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6EQUANEU}},
  note         = {Machine review of arXiv:2507.04123}
}
read the original abstract

This paper presents Edge-based Mixture of Experts (MoE) Collaborative Computing (EMC2), an optimal computing system designed for autonomous vehicles (AVs) that simultaneously achieves low-latency and high-accuracy 3D object detection. Unlike conventional approaches, EMC2 incorporates a scenario-aware MoE architecture specifically optimized for edge platforms. By effectively fusing LiDAR and camera data, the system leverages the complementary strengths of sparse 3D point clouds and dense 2D images to generate robust multimodal representations. To enable this, EMC2 employs an adaptive multimodal data bridge that performs multi-scale preprocessing on sensor inputs, followed by a scenario-aware routing mechanism that dynamically dispatches features to dedicated expert models based on object visibility and distance. In addition, EMC2 integrates joint hardware-software optimizations, including hardware resource utilization optimization and computational graph simplification, to ensure efficient and real-time inference on resource-constrained edge devices. Experiments on open-source benchmarks clearly show the EMC2 advancements as an end-to-end system. On the KITTI dataset, it achieves an average accuracy improvement of 3.58% and a 159.06% inference speedup compared to 15 baseline methods on Jetson platforms, with similar performance gains on the nuScenes dataset, highlighting its capability to advance reliable, real-time 3D object detection tasks for AVs. The official implementation is available at https://github.com/LinshenLiu622/EMC2.

Figures

Figures reproduced from arXiv: 2507.04123 by the authors.

Figure 1
Figure 1. Comparison of 3D object detection models in ac￾curacy and inference efficiency on the KITTI and nuScenes dataset. Most existing methods improve accuracy at the expense of inference efficiency. The proposed EMC2 achieves win-win performance by customizing MoE for autonomous driving. Even with the assistance of low-power machine learning compilation frameworks [6], such systems may still exe￾cute redundant computation… view at source ↗
Figure 2
Figure 2. System Structure of EMC2. The central row illustrates the architecture of EMC2, while the top and bottom rows provide visual examples of core modules and relevant legends. The system integrates five main components: an Adaptive Multimodal Data Bridge (AMDB) for preprocessing raw multimodal input and generating expert-specific features; a Scenario-Adaptive Dispatcher that dispatches suitable experts based on scenario… view at source ↗
Figure 3
Figure 3. Routing Strategy of the Scenario-Adaptive Dispatcher. The Routing criteria is detailed mathematically in Sec. 3.2. first projects proposal regions from the AMDB into BEV space, then extracts 2D features for object localization and classification. It finally maps detection results back to 3D space. Compared to sparse 3D convolutions, its use of 2D operations reduces computation in outdoor ADS scenar￾ios [15], allowin… view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Illustration of Multiscale Pooling and Multimodal Fusion. Proposal regions with image features are first obtained from the 3D proposal space generated by the LiDAR branch of AMDB. After extracting the relevant image features from the cor￾responding image regions and pr…
Figure 6
Figure 6. Figure 6: Illustration of Thread Management. Each solid box represents a module occupying threads. During inference, four boundaries are defined at which preceding threads are terminated, and their thread-specific address space is released upon crossing each boundary. Method Ped…
Figure 7
Figure 7. Figure 7: Distribution of expert activation. During EMC2 in￾ference on the nuScenes validation set, scenes are grouped into a 2D grid based on two metrics: the y-axis indicates the number of AMDB-generated proposals beyond the distance threshold D (i.e., distant objects), and th…
Figure 8
Figure 8. Figure 8: Median Detection Confidence vs. Object Distances. Four configurations are compared on the KITTI validation set: LPE (blue), VEE (peach), APE (rose), and EMC2 (green). More complex experts yield higher confidence but slower inference. By switching experts based on dista…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 43 canonical work pages

  1. [1]

    Haque, Lingjia Tang, and Jason Mars

    Shih-Chieh Lin, Yunqi Zhang, Chang-Hong Hsu, Matt Skach, Md E. Haque, Lingjia Tang, and Jason Mars. The ar- chitectural implications of autonomous driving: Constraints and acceleration. In Proceedings of the Twenty-Third Inter- national Conference on Architectural Support for Program- ming Languages and Operating Systems, ASPLOS ’18, page 751–766, New Yor...

    work page 2018

  2. [2]

    Multimodal end-to-end autonomous driving

    Yi Xiao, Felipe Codevilla, Akhil Gurram, Onay Urfalioglu, and Antonio M L ´opez. Multimodal end-to-end autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 23(1):537–547, 2020. 1

    work page 2020

  3. [3]

    Voxel R-CNN: Towards High Performance Voxel-based 3D Object Detection

    Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. V oxel r-cnn: To- wards high performance voxel-based 3d object detection. arXiv:2012.15712, 2020. 1, 3, 7

    work page Pith review arXiv 2012

  4. [4]

    Vlm2scene: Self-supervised image-text-lidar learning with foundation models for autonomous driving scene understanding

    Guibiao Liao, Jiankun Li, and Xiaoqing Ye. Vlm2scene: Self-supervised image-text-lidar learning with foundation models for autonomous driving scene understanding. InPro- ceedings of the AAAI Conference on Artificial Intelligence , volume 38, pages 3351–3359, 2024. 1

    work page 2024

  5. [5]

    Optimizing fpga-based accelerator design for deep convolutional neural networks

    Chen Zhang, Peng Li, Guangyu Sun, Yijin Guan, Bingjun Xiao, and Jason Cong. Optimizing fpga-based accelerator design for deep convolutional neural networks. In Proceed- ings of the 2015 ACM/SIGDA international symposium on field-programmable gate arrays, pages 161–170, 2015. 1

    work page 2015

  6. [6]

    {TVM}: An automated {End- to-End} optimizing compiler for deep learning

    Tianqi Chen, Thierry Moreau, Ziheng Jiang, Lianmin Zheng, Eddie Yan, Haichen Shen, Meghan Cowan, Leyuan Wang, Yuwei Hu, Luis Ceze, et al. {TVM}: An automated {End- to-End} optimizing compiler for deep learning. In 13th USENIX Symposium on Operating Systems Design and Im- plementation (OSDI 18), pages 578–594, 2018. 2, 3

    work page 2018

  7. [7]

    Ace: Ally complementary experts for solving long-tailed recognition in one-shot

    Jiarui Cai, Yizhou Wang, and Jenq-Neng Hwang. Ace: Ally complementary experts for solving long-tailed recognition in one-shot. In Proceedings of the IEEE/CVF international conference on computer vision, pages 112–121, 2021. 2

    work page 2021

  8. [8]

    Uni- moe: Scaling unified multimodal llms with mixture of ex- perts

    Yunxin Li, Shenyuan Jiang, Baotian Hu, Longyue Wang, Wanqi Zhong, Wenhan Luo, Lin Ma, and Min Zhang. Uni- moe: Scaling unified multimodal llms with mixture of ex- perts. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2

    work page 2025

Show all 52 references
  1. [9]

    Cumo: Scaling multimodal llm with co-upcycled mixture- of-experts

    Jiachen Li, Xinyao Wang, Sijie Zhu, Chia-Wen Kuo, Lu Xu, Fan Chen, Jitesh Jain, Humphrey Shi, and Longyin Wen. Cumo: Scaling multimodal llm with co-upcycled mixture- of-experts. Advances in Neural Information Processing Sys- tems, 37:131224–131246, 2024. 2

    2024

  2. [10]

    A survey on mixture of experts.arXiv preprint arXiv:2407.06204, 2024

    Weilin Cai, Juyong Jiang, Fan Wang, Jing Tang, Sunghun Kim, and Jiayi Huang. A survey on mixture of experts.arXiv preprint arXiv:2407.06204, 2024. 2

    2024 arXiv

  3. [11]

    Time-moe: Billion-scale time series foundation models with mixture of experts

    Xiaoming Shi, Shiyu Wang, Yuqi Nie, Dianqi Li, Zhou Ye, Qingsong Wen, and Ming Jin. Time-moe: Billion-scale time series foundation models with mixture of experts. arXiv preprint arXiv:2409.16040, 2024. 2

    2024 arXiv

  4. [12]

    Fastmoe: A fast mixture-of-expert train- ing system

    Jiaao He, Jiezhong Qiu, Aohan Zeng, Zhilin Yang, Jidong Zhai, and Jie Tang. Fastmoe: A fast mixture-of-expert train- ing system. arXiv preprint arXiv:2103.13262, 2021. 2

    2021 arXiv

  5. [13]

    Qi, Xinlei Chen, Or Litany, and Leonidas J

    Charles R. Qi, Xinlei Chen, Or Litany, and Leonidas J. Guibas. Imvotenet: Boosting 3d object detection in point clouds with image votes. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4403–4412, 2020. 2

    2020

  6. [14]

    From points to parts: 3d object detec- tion from point cloud with part-aware and part-aggregation network

    Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detec- tion from point cloud with part-aware and part-aggregation network. arXiv preprint arXiv:1907.03670, 2019. 3

    1907 arXiv

  7. [15]

    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. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12689–12697, 2019. 2, 3, 4, 7

    2019

  8. [16]

    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. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 4490–4499, 2018. 3

    2018

  9. [17]

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

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

    2020

  10. [18]

    Center- based 3d object detection and tracking

    Tianwei Yin, Xingyi Zhou, and Philipp Kr ¨ahenb¨uhl. Center- based 3d object detection and tracking. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11779–11788, 2021. 3

    2021

  11. [19]

    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. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11037–11045, 2020. 3

    2020

  12. [20]

    Full waveform lidar for adverse weather conditions

    Andrew M Wallace, Abderrahim Halimi, and Gerald S Buller. Full waveform lidar for adverse weather conditions. IEEE transactions on vehicular technology , 69(7):7064– 7077, 2020. 3

    2020

  13. [21]

    What happens for a tof lidar in fog? IEEE Trans- actions on Intelligent Transportation Systems, 22(11):6670– 6681, 2020

    You Li, Pierre Duthon, Michele Colomb, and Javier Ibanez- Guzman. What happens for a tof lidar in fog? IEEE Trans- actions on Intelligent Transportation Systems, 22(11):6670– 6681, 2020. 3

    2020

  14. [22]

    Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera

    Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In 2019 international conference on robotics and automation (ICRA), pages 3288–3295. IEEE, 2019. 3

    2019

  15. [23]

    Ep- net: Enhancing point features with image semantics for 3d object detection

    Tengteng Huang, Zhe Liu, Xiwu Chen, and Xiang Bai. Ep- net: Enhancing point features with image semantics for 3d object detection. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020 , pages 35–52, Cham, 2020. Springer Int...

    2020

  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]

    Unifying voxel-based representation with transformer for 3d object detection

    Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. In Advances in Neural Information Processing Systems, 2022. 3

    2022

  18. [26]

    Ppf-det: Point-pixel fusion for multi-modal 3d object detection

    Guotao Xie, Zhiyuan Chen, Ming Gao, Manjiang Hu, and Xiaohui Qin. Ppf-det: Point-pixel fusion for multi-modal 3d object detection. IEEE Transactions on Intelligent Trans- portation Systems, 25(6):5598–5611, 2024. 3

    2024

  19. [27]

    Robofusion: To- wards robust multi-modal 3d obiect detection via sam

    Ziying Song, Guoxing Zhang, Lin Liu, Lei Yang, Shaoqing Xu, Caiyan Jia, Feiyang Jia, and Li Wang. Robofusion: To- wards robust multi-modal 3d obiect detection via sam. arXiv preprint arXiv:2401.03907, 2024. 3

    2024 arXiv

  20. [28]

    Machine learn- ing in compiler optimization

    Zheng Wang and Michael O’Boyle. Machine learn- ing in compiler optimization. Proceedings of the IEEE , 106(11):1879–1901, 2018. 3

    1901

  21. [29]

    Au- tomatic feature generation for machine learning–based opti- mising compilation

    Hugh Leather, Edwin Bonilla, and Michael O’boyle. Au- tomatic feature generation for machine learning–based opti- mising compilation. ACM Transactions on Architecture and Code Optimization (TACO), 11(1):1–32, 2014. 3

    2014

  22. [30]

    NVIDIA TensorRT Developer Guide,

    NVIDIA Corporation. NVIDIA TensorRT Developer Guide,

  23. [31]

    XLA: Compiling Machine Learning for Peak Performance, 2020

    Amit Sabne. XLA: Compiling Machine Learning for Peak Performance, 2020. 3

    2020

  24. [32]

    Onnx: Open neural network exchange

    ONNX Community. Onnx: Open neural network exchange. https://github.com/onnx/onnx, 2024. Accessed: YYYY-MM-DD. 3

    2024

  25. [33]

    U- net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image segmen- tation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, pa...

    2015

  26. [34]

    Adaptive mixtures of local experts.Neu- ral computation, 3(1):79–87, 1991

    Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts.Neu- ral computation, 3(1):79–87, 1991. 5

    1991

  27. [35]

    Second: Sparsely em- bedded convolutional detection

    Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely em- bedded convolutional detection. Sensors, 18(10), 2018. 6, 7

    2018

  28. [36]

    Communication complexity and parallel computing

    Juraj Hromkovi ˇc. Communication complexity and parallel computing. Springer Science & Business Media, 2013. 6

    2013

  29. [37]

    Are we ready for autonomous driving? the kitti vision benchmark suite

    Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3354–3361, 2012. 6

    2012

  30. [38]

    nuscenes: A multi- modal dataset for autonomous driving

    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 Proceedings of the IEEE/CVF conference on computer vision and patte...

    2020

  31. [39]

    Introducing kolmogorov–smirnov tests under uncertainty: an application to radioactive data

    Muhammad Aslam. Introducing kolmogorov–smirnov tests under uncertainty: an application to radioactive data. ACS omega, 5(1):914–917, 2019

    2019

  32. [40]

    M3detr: Multi- representation, multi-scale, mutual-relation 3d object detec- tion with transformers

    Tianrui Guan, Jun Wang, Shiyi Lan, Rohan Chandra, Zux- uan Wu, Larry Davis, and Dinesh Manocha. M3detr: Multi- representation, multi-scale, mutual-relation 3d object detec- tion with transformers. In Proceedings of the IEEE/CVF winter conference on applications of computer vis...

    2022

  33. [41]

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

    Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point- voxel feature set abstraction for 3d object detection. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10529–10538, 2020. 7

    2020

  34. [42]

    Gle-net: A global and local ensemble network for aerial object detection

    Jiajia Liao, Yujun Liu, Yingchao Piao, Jinhe Su, Guorong Cai, and Yundong Wu. Gle-net: A global and local ensemble network for aerial object detection. International Journal of Computational Intelligence Systems, 15(1):2, 2022. 7

    2022

  35. [43]

    From points to parts: 3d object detection from point cloud with part-aware and part-aggregation net- work

    Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation net- work. IEEE transactions on pattern analysis and machine intelligence, 43(8):2647–2664, 2020. 7

    2020

  36. [44]

    3onet: 3d detector for occluded object under obstructed conditions

    Hiep Anh Hoang and Myungsik Yoo. 3onet: 3d detector for occluded object under obstructed conditions. IEEE Sensors Journal, 2023. 7

    2023

  37. [45]

    Logonet: Towards accurate 3d object detection with local-to-global cross-modal fusion

    Xin Li, Tao Ma, Yuenan Hou, Botian Shi, Yuchen Yang, Youquan Liu, Xingjiao Wu, Qin Chen, Yikang Li, Yu Qiao, et al. Logonet: Towards accurate 3d object detection with local-to-global cross-modal fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern R...

    2023

  38. [46]

    V oxel field fusion for 3d object detection

    Yanwei Li, Xiaojuan Qi, Yukang Chen, Liwei Wang, Zem- ing Li, Jian Sun, and Jiaya Jia. V oxel field fusion for 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 7

    2022

  39. [47]

    Cat-det: Con- trastively augmented transformer for multi-modal 3d object detection

    Yanan Zhang, Jiaxin Chen, and Di Huang. Cat-det: Con- trastively augmented transformer for multi-modal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 908–917,

  40. [48]

    Transformation-equivariant 3d object detection for autonomous driving

    Hai Wu, Chenglu Wen, Wei Li, Xin Li, Ruigang Yang, and Cheng Wang. Transformation-equivariant 3d object detection for autonomous driving. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2795–2802, 2023. 7

    2023

  41. [49]

    Nils Lubbe, Yi Wu, and Hanna Jeppsson. Safe speeds: fa- tality and injury risks of pedestrians, cyclists, motorcyclists, and car drivers impacting the front of another passenger car as a function of closing speed and age. Traffic safety re- search, 2:000006–000006, 2022. 7

    2022

  42. [50]

    Focal- former3d: focusing on hard instance for 3d object detection

    Yilun Chen, Zhiding Yu, Yukang Chen, Shiyi Lan, An- ima Anandkumar, Jiaya Jia, and Jose M Alvarez. Focal- former3d: focusing on hard instance for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8394–8405, 2023. 7

    2023

  43. [51]

    Rus, and Song Han

    Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L. Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation. In 2023 IEEE International Conference on Robotics and Automation (ICRA) , pages 2774–2781, 2023. 7 10

    2023

  44. [2024]

    Accessed: 2025-02-13. 3

    2025

Pith tools

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