Pith. sign in

REVIEW 3 major objections 4 minor 71 references

AWML: An Open-Source ML-based Robotics Perception Framework to Deploy for ROS-based Autonomous Driving Software

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read AWML is an open-source MLOps framework for ROS-based autonomous driving that carries perception models from training through ONNX and TensorRT deployment into Autoware, then feeds deployment data back through pseudo-labeling and…

desk verdict Useful engineering framework for Autoware MLOps, but the benchmark tables are unverifiable until the paper says which models were trained on test data under the product-release procedure. read the letter →

arxiv 2506.00645 v1 pith:4XSXZ6AR submitted 2025-05-31 cs.RO cs.SE

Satoshi Tanaka , Samrat Thapa , Kok Seang Tan , Amadeusz Szymko , Lobos Kenzo , Koji Minoda , Shintaro Tomie , Kotaro Uetake
show 3 more authors
Guolong Zhang Isamu Yamashita Takamasa Horibe
This is my paper
classification cs.ROcs.SE
keywords MLOpsautonomousdrivingroboticsperceptionactivelearningpseudo-labeling3DobjectdetectionAutowareROS2
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

AWML is an open-source MLOps framework whose central claim is that a single pipeline can carry a robotics perception model from training to deployment to retraining: models are trained on a unified dataset format, exported to ONNX and TensorRT, run inside the ROS-based Autoware stack, and the same unlabeled driving data is fed back through pseudo-labeling to produce new training data. The framework organizes models into a five-stage hierarchy (pretrain, base, product, project, and offline) and versioning that tells operators when a model update is compatible with existing ROS packages. The concrete evidence that the loop pays off is a fine-tuning experiment on a bus dataset, where the product model improves mean average precision from 64.5 to 65.4 and bus AP from 86.6 to 90.4. A reader should care because the paper addresses a known gap: computer-vision models are typically developed in Python and evaluated offline, while robots run in C++ under real-time constraints, and AWML aims to be the bridge between those two worlds.

What carries the argument

The central object is the AWML pipeline: a set of scripts and configuration files layered on the open-source 2D and 3D detection toolboxes, tied together by the T4dataset, a unified dataset format that stores database, use-case, non-annotated, and pseudo-labeled data in one structure. The mechanism that carries the argument is the five-stage model hierarchy (pretrain, base, product, project, and offline models for labeling), together with a product-release procedure that fine-tunes the final deployment model on the full dataset once a configuration has been selected. Versioning strings such as 'CenterPoint bus/1.2.3-odaiba.2' map every model to the ROS parameters it depends on, which is what makes deployment and retraining consistent within Autoware.

What would settle it

Check the training configuration for each row in Tables 2 through 7 to see whether the checkpoint evaluated on the test split was ever trained on that split; then rerun the Table 7 comparison with the product model fine-tuned on the training split only and evaluated on the untouched test split. If the mAP gain from 64.5 to 65.4 and the bus AP gain from 86.6 to 90.4 disappear or reverse under that protocol, the paper's central quantitative claim is not supported.

Watch

Extended reading notes

Core claim

The paper claims that MLOps for robotics can be made practical by building a training and deployment framework directly on the detection toolboxes the vision community already uses, and pairing that with an active-learning pipeline that treats unlabeled rosbag data as a reusable resource. Models are trained as broad base models on the paper's unified T4dataset, fine-tuned into product models matched to specific vehicle sensor configurations, and supplemented by offline models that generate pseudo-labels for the next training round. Deployment exports trained weights to ONNX and TensorRT so they execute inside Autoware, and the same dataset format underlies annotation, pseudo-labeling, and evaluation. In the experiments, the four 3D detectors evaluated reach mAP values in the mid-to-high 60s on the internal taxi and bus datasets, and the reported fine-tuning gain on the bus dataset is mAP 64.5 to 65.4 with bus AP 86.6 to 90.4, which the paper presents as evidence that product-specific fine-tuning works in practice.

Load-bearing premise

The benchmark tables are assumed to report genuine generalization on held-out test data, even though Section 4.1 says the models were trained on all available data and Section 3.7 describes a product-release procedure that deliberately retrains on the train, validation, and test splits combined.

Editorial extensions

If this is right

  • A team using AWML can ship a perception update without writing ROS glue code, because training, ONNX export, TensorRT compilation, and Autoware parameter files are handled by the pipeline.
  • The active-learning loop turns raw driving logs into pseudo-labeled T4dataset entries, so product models can be updated with deployment data that never required manual annotation.
  • Product models fine-tuned on sensor-matched vehicle data outperform the general base model on their target platform, as the reported bus mAP and AP gains show.
  • The versioning scheme tells an Autoware operator whether a model update is a drop-in change (same major version) or requires a ROS package review (major version bump).
  • TensorRT deployment reduces reported inference latency by roughly an order of magnitude, for example 558.2 ms to 29.1 ms for BEVFusion-L, making the models usable in real-time Autoware modules.

Reading between the lines

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

  • The paper does not quantify the annotation cost saved by its pseudo-labeling loop; a testable extension is to compare mAP per annotated frame with and without the active-learning pipeline.
  • Because the paper describes both training 'by all dataset' and a product-release procedure that retrains on train, validation, and test splits combined, the numerical comparisons are best read as engineering validation of the pipeline rather than as head-to-head evidence about which detection architecture is best.
  • The T4dataset's compatibility with a public dataset format suggests that AWML-trained models could be evaluated with the same tooling used in the research literature, which would lower the cost of research-to-production transfer if those evaluation scripts are released.
  • The five-stage model hierarchy implies a fleet-oriented deployment model: one shared base model with per-vehicle fine-tunes, while project-specific data stays private and outside the shared training set.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. This paper presents AWML, an open-source MLOps framework for ROS 2 and Autoware-based autonomous driving perception. It describes a training-to-deployment pipeline built on MMDetection and MMDetection3D, a T4dataset format derived from nuScenes, a model versioning scheme covering pretrain, base, product, and project models, and an active-learning workflow based on pseudo-labeling. The experimental section benchmarks CenterPoint, PillarNeSt, BEVFusion-L, and BEVFusion-CL on internal taxi and bus datasets, reports inference latencies in PyTorch and TensorRT with ROS 2, and claims that product fine-tuning improves bus detection accuracy.

Significance. As a systems contribution, AWML addresses a real gap by giving computer-vision researchers a path to deploy models in ROS-based autonomous driving stacks and by integrating versioning, fine-tuning, and pseudo-labeling into one workflow. The reported TensorRT latency figures show large speedups over PyTorch, and the explicit documentation of release policies is useful for practitioners. However, the quantitative accuracy claims are compromised by unresolved train/test split ambiguity, and the active-learning pipeline is not experimentally validated. If the split issue is fixed and the benchmark tables are recomputed under a standard protocol, the framework would be a solid contribution to the robotics MLOps literature.

major comments (3)
  1. [§4.2 / Table 7 / §3.7] The central fine-tuning claim is uninterpretable as written. Section 3.7 explicitly defines a product-release model that is fine-tuned on the full dataset including the test split, while Table 7 labels the fine-tuned row only as "fine-tuning as product model." A "product model" could mean the step-1 model trained on the training split or the step-3 product-release model trained on the test split. The reported mAP gain from 64.5 to 65.4 and bus AP gain from 86.6 to 90.4 therefore cannot be read as evidence of generalization unless the row is explicitly identified as using standard splits. Please state for every row in Tables 2-4, 6, and 7 which data splits were used for training, and recompute the comparison under the standard train/val/test protocol.
  2. [§4.1 / Tables 2-4] Section 4.1 says "We train 3D object detection models by all dataset with Taxi dataset and Bus dataset." If "all dataset" includes the evaluation frames, then the benchmark numbers in Tables 2-4 are not estimates of generalization and cannot support method comparisons. The paper needs to state explicitly that all benchmark models were trained only on the training split, and that the product-release procedure of Section 3.7 was applied only to explicitly labeled rows, if any.
  3. [§3.1 / §4.2 / §4.3] The abstract and Section 1 list active learning with auto-labeling, semi-auto-labeling, and data mining as core contributions, but the experiments contain no evaluation of this pipeline. There are no pseudo-label quality metrics, no report of how many frames were auto-labeled, no comparison of model performance with and without pseudo-labeled data, and no sensitivity analysis for the confidence thresholds mentioned in Section 3.1. The claim that this pipeline enables more efficient and adaptive model training is therefore not supported by any measurement in the paper.
minor comments (4)
  1. [Throughout] The paper describes AWML as open-source but does not provide a public repository URL; please include the link to the released code.
  2. [References [3] and [7]] Reference [7] is titled "Bevformer-tensorrt" but the cited project is YOLOv4-Darknet-TensorRT; the reference label and the in-text citation in Table 1 should be corrected to match the actual project.
  3. [Tables 5 and 10] Latency measurements are reported as single numbers without error bars, number of runs, warmup procedure, or GPU-clock details; please state the measurement methodology or provide at least a few repeated runs to support the real-time claims.
  4. [Title and headers] The PDF header spells the framework name as "A WML" while the title uses "AWML"; please harmonize the formatting.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: AWML is an engineering/MLOps framework; benchmark results are empirical and the product-release retraining is disclosed rather than disguised as prediction.

full rationale

AWML does not present a derivation chain in which an output quantity is defined by an input quantity. The pipeline components (training, deployment, pseudo-labeling, fine-tuning) are engineering workflows, not equations. The only potentially self-referential element is the product-release model in Sec. 3.7, which is explicitly described as fine-tuning on the full dataset including train, val, and test splits to maximize deployment performance; this is disclosed and is not presented as a generalization benchmark. Table 7 labels its fine-tuned row as "fine-tuning as product model," and Sec. 3.7 defines the product model as the step-1 model trained on the training split, so the reported w/FT gain is not shown to be forced by construction. The pseudo-labeling loop uses the authors' offline models, but self-training is an empirical process, not a definitional equivalence. The only self-citation is the WebAuto system [58] used to manage T4dataset IDs in the appendix; it is not load-bearing for any claim. The unresolved ambiguity about whether Tables 2-4 report standard splits or product-release retraining is a data-hygiene/soundness concern, not circularity, and cannot by itself establish that any result reduces to its own input.

Assumptions & free parameters 2 free parameters · 5 assumptions · 1 invented entities

The paper introduces no new physical entities. Its central claims rest on the correctness of the underlying ML libraries, the Autoware and ROS 2 integration, and the quality of private annotations and pseudo-labels. Free parameters are limited to user-adjustable thresholds and evaluation-range choices; no fitted constants are used.

free parameters (2)
  • Pseudo-label confidence threshold = not specified (adjustable in choose_annotation.py)
    This threshold filters raw pseudo-labels before training (Sec. 3.1). No value or sensitivity analysis is given, and it directly determines the quality of auto-labeled training data.
  • Evaluation range = 120 m
    The evaluation range is set to 120 m for T4dataset, whereas nuScenes uses 50 m (Sec. 4.2). This choice affects all mAP values and complicates comparison with prior published results.
assumptions (5)
  • domain assumption Correctness of MMDetection and MMDetection3D implementations is assumed.
    AWML is built directly on these toolkits and uses their training, evaluation, and export routines (Sec. 3.1). If they contain bugs or configuration mismatches, the reported benchmark numbers change.
  • domain assumption T4dataset annotations are accurate and unbiased.
    All experiments depend on annotations of the private dataset; no inter-annotator agreement, annotation QA, or public release is described (Sec. 4.1, App. 6.3).
  • domain assumption Reported test benchmarks used uncontaminated train/test splits.
    Sec. 4.1 says models are trained 'by all dataset' while Tables 2-4 report test-set results, and Sec. 3.7 describes training product-release models on train plus val plus test. The paper never reconciles these statements.
  • ad hoc to paper Pseudo-labels generated by offline models are reliable enough for training.
    The active-learning pipeline (Secs. 3.1, 3.2, 3.4) trains pretrain and project models on pseudo T4datasets without any validation of pseudo-label accuracy.
  • domain assumption Autoware and ROS 2 deployment packages correctly load the exported models.
    Deployment results (Tables 5 and 10) assume that the Autoware ML package and ROS 2 wrapper correctly execute the exported ONNX and TensorRT models.
invented entities (1)
  • T4dataset
    purpose: Unified in-house dataset format based on nuScenes, used to train and evaluate models across open and private projects.
    The format is only described textually; the dataset itself is private and never released, so the new resource that anchors every experiment cannot be independently checked.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AWML: An Open-Source ML-based Robotics Perception Framework to Deploy for ROS-based Autonomous Driving Software." pith.science (2026). https://pith.science/paper/4XSXZ6AR

@misc{pith2026250600645,
  author       = {Pith},
  title        = {Pith review of: AWML: An Open-Source ML-based Robotics Perception Framework to Deploy for ROS-based Autonomous Driving Software},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4XSXZ6AR}},
  note         = {Machine review of arXiv:2506.00645}
}
read the original abstract

In recent years, machine learning technologies have played an important role in robotics, particularly in the development of autonomous robots and self-driving vehicles. As the industry matures, robotics frameworks like ROS 2 have been developed and provides a broad range of applications from research to production. In this work, we introduce AWML, a framework designed to support MLOps for robotics. AWML provides a machine learning infrastructure for autonomous driving, supporting not only the deployment of trained models to robotic systems, but also an active learning pipeline that incorporates auto-labeling, semi-auto-labeling, and data mining techniques.

Figures

Figures reproduced from arXiv: 2506.00645 by the authors.

Figure 1
Figure 1. The framework of AWML and ongoing optimization. To address this complexity, the concept of Machine Learning Operations (MLOps) has gained traction. MLOps practices aim to ensure that ma￾chine learning models are reproducible, stable, and main￾tainable throughout their lifecycle—from initial develop￾ment to deployment in production systems. However, in￾tegrating MLOps across both computer vision and robotics domains … view at source ↗
Figure 2
Figure 2. The pipeline of AWML. ware deployment—play a key role in enhancing the per￾formance of online machine learning models by leverag￾ing offline training. The process begins by generating an info file from an unlabeled T4dataset for both 2D and 3D data using the pseudo label.py script (6). This is fol￾lowed by the creation of a pseudo-labeled T4dataset via create pseudo t4dataset.py (7), and then refinement of an￾notati… view at source ↗
Figure 3
Figure 3. Model examples and T4dataset. figuration of the target application. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Model types for fine-tuning pipelines in LiDAR-only 3D object detection. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Model types for fine-tuning pipelines in Camera-LiDAR-Fusion 3D object detection. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Usage of model. CenterPoint base/1.0 Base model CenterPoint base/1.1 Product model CenterPoint bus/1.0.1 CenterPoint bus/1.0.0 CenterPoint bus/1.0.1- {project name}.1 Project model Pretrain model CenterPoint pretrain/20241203 Fine-tuning CenterPoint bus/1.0.1- {project…
Figure 7
Figure 7. Figure 7: Model release strategy. ber 3rd, 2024. To update the version of pretrain model from from “pretrain/{date}” to “pretrain/{next date}.” We prepare the pretrain model using a pseudo T4dataset to enhance generalization performance. Pseudo T4dataset contains various vehicle…
Figure 8
Figure 8. Figure 8: Introduction of new algorithm. Base model Train dataset 1. Fine-tuning for product model Product model Train dataset Val dataset Test dataset Dataset using for training Train dataset 3. Fine-tuning for product-release model Product-release model Val dataset Test datase…
Figure 9
Figure 9. Figure 9: Fine-tuning strategy. BEVFusion-L is LiDAR-only model using sparse convo￾lution encoder. BEVFusion-CL is Camera-LiDAR fusion model. In Sec. 4.3, we evaluate 2D detection for traffic light us￾ing T4dataset for traffic light recognition. In this exper￾iment, we use YOLOX…
Figure 11
Figure 11. Figure 11: The structure of autoware ml projects BEVFusion CenterPoint MobileNetv2 YOLOX opt [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: The structure of projects tools classification2d detection2d detection3d [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]
Figure 10
Figure 10. Figure 10: The whole structure of AWML autoware ml configs detection3d XX1.py X2.py detection2d XX1.py X2.py t4dataset db jpntaxi v1.yaml db jpntaxi v2.yaml db jpntaxi v3.yaml tlr v1 0.yaml [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 14
Figure 14. Figure 14: The structure of T4dataset {root} AWML models info auto label [PITH_FULL_IMAGE:figures/full_fig_p016_14.png]
Figure 15
Figure 15. Figure 15: The structure of model zoo . ” T4dataset WebAuto version” is the version of T4dataset itself. If we fix annotation or sensor data, we update this version. When we make a T4dataset, we start from version 0. t4dev-kit. The software of “t4dev-kit” provide the API to use …
Figure 16
Figure 16. Figure 16: The structure of model data 3 [PITH_FULL_IMAGE:figures/full_fig_p017_16.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

71 extracted references · 49 canonical work pages

  1. [1]

    com / ApolloAuto / apollo

    Apollo.https : / / github . com / ApolloAuto / apollo. 3, 4

    work page

  2. [2]

    com / autowarefoundation / autoware

    autoware.universe.https : / / github . com / autowarefoundation / autoware . universe. 3, 4

    work page

  3. [3]

    com / DerryHub/BEVFormer_tensorrt

    Bevformer-tensorrt.https : / / github . com / DerryHub/BEVFormer_tensorrt. 3, 4

    work page

  4. [4]

    Dl4agx.https://github.com/NVIDIA/DL4AGX. 3

    work page

  5. [5]

    Nvidia-ai-iot.https://github.com/NVIDIA- AI- IOT/Lidar_AI_Solution. 3, 4

    work page

  6. [6]

    com / ksm26 / ROS - based - 3D - detection - Tracking

    Ros-based-3d-detection-tracking.https : / / github . com / ksm26 / ROS - based - 3D - detection - Tracking. 3, 4

    work page

  7. [7]

    com / laitathei / YOLOv4 - Darknet - TensorRT

    Bevformer-tensorrt.https : / / github . com / laitathei / YOLOv4 - Darknet - TensorRT. 3, 4

    work page

  8. [8]

    mmdetection-ros.https://github.com/jcuic5/ mmdetection-ros. 3, 4

    work page

Show all 71 references
  1. [9]

    Open-source software toolkit for optimizing and deploy- ing deep learning models.https://github.com/ openvinotoolkit/openvino. 3

  2. [10]

    Parallel distributed deep learning: Machine learning frame- work from industrial practice.https://github.com/ PaddlePaddle/Paddle. 3

  3. [11]

    com / intel/ros2_openvino_toolkit

    ros2 openvino toolkit.https : / / github . com / intel/ros2_openvino_toolkit. 3, 4

  4. [12]

    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. InECCV Workshops, 2018. 2

    2018

  5. [13]

    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, 3

    2022

  6. [14]

    Yolo-z: Improving small object detection in yolov5 for autonomous vehicles.ArXiv, abs/2112.11798,

    Aduen Benjumea, Izzedin Teeti, Fabio Cuzzolin, and An- drew Bradley. Yolo-z: Improving small object detection in yolov5 for autonomous vehicles.ArXiv, abs/2112.11798,

    arXiv

  7. [15]

    Yolov4: Optimal speed and accuracy of object detection, 2020

    Alexey Bochkovskiy, Chien-Yao Wang, and Hong- Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection, 2020. 3, 4

    2020

  8. [16]

    Yolobile: Real- time object detection on mobile devices via compression- compilation co-design.ArXiv, abs/2009.05697, 2020

    Yuxuan Cai, Hongjia Li, Geng Yuan, Wei Niu, Yanyu Li, Xulong Tang, Bin Ren, and Yanzhi Wang. Yolobile: Real- time object detection on mobile devices via compression- compilation co-design.ArXiv, abs/2009.05697, 2020. 2

    2009 arXiv

  9. [17]

    MMDetection: Open mmlab detection toolbox and benchmark.arXiv preprint arXiv:1906.07155, 2019

    Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tian- heng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyan...

    1906 arXiv

  10. [18]

    Zhang, Jian Cheng, and Jian Sun

    Qiang Chen, Yingming Wang, Tong Yang, X. Zhang, Jian Cheng, and Jian Sun. You only look one-level feature. 12 2021 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 13034–13043, 2021. 2

    2021

  11. [19]

    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

  12. [20]

    Multi-view 3d object detection network for autonomous driving.2017 IEEE Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 6526–6534, 2016

    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

  13. [21]

    MMDetection3D: Open- MMLab next-generation platform for general 3D object detection.https://github.com/open- mmlab/ mmdetection3d, 2020

    MMDetection3D Contributors. MMDetection3D: Open- MMLab next-generation platform for general 3D object detection.https://github.com/open- mmlab/ mmdetection3d, 2020. 2, 3

    2020

  14. [22]

    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

  15. [23]

    Yolox: Exceeding yolo series in 2021.ArXiv, abs/2107.08430, 2021

    Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021.ArXiv, abs/2107.08430, 2021. 2, 3, 4, 10

    2021 arXiv

  16. [24]

    Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco An- dreetto, and Hartwig Adam

    Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco An- dreetto, and Hartwig Adam. Mobilenets: Efficient convolu- tional neural networks for mobile vision applications.ArXiv, abs/1704.04861, 2017. 2

    2017 arXiv

  17. [25]

    Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, Quoc V

    Andrew G. Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, Quoc V . Le, and Hartwig Adam. Searching for mobilenetv3.2019 IEEE/CVF Interna- tional Conference on Computer Vision (ICCV), pages 1314– 1...

    2019

  18. [26]

    Bevdet: High-performance multi-camera 3d object detection in bird-eye-view.ArXiv, abs/2112.11790, 2021

    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. 2

    2021 arXiv

  19. [27]

    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. InComputer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29 – October 4, 2024, Proceedings, Part XXII, page 439–455...

    2024

  20. [28]

    Multi-modal sensor fusion for auto driving perception: A survey.ArXiv, abs/2202.02703, 2022

    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

  21. [29]

    Adriver-i: A general world model for autonomous driving

    Fan Jia, Weixin Mao, Yingfei Liu, Yucheng Zhao, Yuqing Wen, Chi Zhang, Xiangyu Zhang, and Tiancai Wang. Adriver-i: A general world model for autonomous driving. arXiv preprint arXiv:2311.13549, 2023. 9

    2023 arXiv

  22. [30]

    Qi, Yin Zhou, and Drago Anguelov

    Chiyu Max Jiang, Mahyar Najibi, C. Qi, Yin Zhou, and Drago Anguelov. Improving the intra-class long-tail in 3d detection via rare example mining.ArXiv, abs/2210.08375,

    arXiv

  23. [31]

    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. InThirty-Eighth AAAI Conference on Artificial Intelli- gence, AAAI 2024, pages 2561–2569. AAAI Press...

    2024

  24. [32]

    Autoware on board: Enabling autonomous vehicles with embedded systems

    Shinpei Kato, Shota Tokunaga, Yuya Maruyama, Seiya Maeda, Manato Hirabayashi, Yuki Kitsukawa, Abraham Monrroy, Tomohito Ando, Yusuke Fujii, and Takuya Azumi. Autoware on board: Enabling autonomous vehicles with embedded systems. In2018 ACM/IEEE 9th International Conference on ...

    2018

  25. [33]

    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, 3, 9

    2019

  26. [34]

    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

  27. [35]

    Bevdepth: Acquisition of reliable depth for multi-view 3d object detection.ArXiv, abs/2206.10092, 2022

    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. 2

    2022 arXiv

  28. [36]

    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. In2022 IEEE/CVF Conference on Computer V...

  29. [37]

    Fully sparse fusion for 3d object detection.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    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

  30. [38]

    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. 2, 3, 4

    2022

  31. [39]

    Bevfusion: A simple and robust lidar-camera fusion framework.ArXiv, abs/2205.13790, 2022

    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

  32. [40]

    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. InIEEE International Conference on Robotics and Automation (ICRA), 2023. 2, 3, 9

    2023

  33. [41]

    Exploring active 3d object detection from a generalization perspective.ArXiv, abs/2301.09249,

    Yadan Luo, Zhuoxiao Chen, Zijian Wang, Xin Yu, Zi Huang, and Mahsa Baktash. Exploring active 3d object detection from a generalization perspective.ArXiv, abs/2301.09249,

    arXiv

  34. [42]

    Robot operating system 2: Design, architecture, and uses in the wild.Science Robotics, 7(66):eabm6074, 2022

    Steven Macenski, Tully Foote, Brian Gerkey, Chris Lalancette, and William Woodall. Robot operating system 2: Design, architecture, and uses in the wild.Science Robotics, 7(66):eabm6074, 2022. 3

    2022

  35. [43]

    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

  36. [44]

    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. InComputer Vision – ECCV 2020: 16th Euro- pean Conference, Glasgow, UK, August 23–28, 2020, Pro- ceedings, Part XIV, page 194–210, Berlin, Heidelb...

    2020

  37. [45]

    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

  38. [46]

    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. InProceedings of the 31st International Conference on Neural Information Processing Systems, page 5105–5114, Red Hook, NY , USA, 2017. Curran...

    2017

  39. [47]

    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. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 918–927,

  40. [48]

    Ros: an open-source robot operating sys- tem

    Morgan Quigley. Ros: an open-source robot operating sys- tem. InIEEE International Conference on Robotics and Au- tomation, 2009. 1, 3

    2009

  41. [49]

    Yolov3: An incremental improvement.ArXiv, abs/1804.02767, 2018

    Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement.ArXiv, abs/1804.02767, 2018. 2

    2018 arXiv

  42. [50]

    Girshick, and Ali Farhadi

    Joseph Redmon, Santosh Kumar Divvala, Ross B. Girshick, and Ali Farhadi. You only look once: Unified, real-time ob- ject detection.2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779–788, 2015. 2

    2016

  43. [51]

    Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen

    Mark Sandler, Andrew G. Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks.2018 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, pages 4510–4520, 2018. 2, 10

    2018

  44. [52]

    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

    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

  45. [53]

    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

  46. [54]

    Mingxing Tan and Quoc V . Le. Efficientnet: Rethinking model scaling for convolutional neural networks.ArXiv, abs/1905.11946, 2019. 2

    1905 arXiv

  47. [55]

    Mingxing Tan and Quoc V . Le. Efficientnetv2: Smaller mod- els and faster training. InInternational Conference on Ma- chine Learning, 2021. 2

    2021

  48. [56]

    Openpcdet: An open- source toolbox for 3d object detection from point clouds

    OpenPCDet Development Team. Openpcdet: An open- source toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet,

  49. [57]

    Fcos: Fully convolutional one-stage object detection.2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9626–9635, 2019

    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. 2

    2019

  50. [58]

    Inc. TIER IV . Webauto.https://web.auto/. 1

  51. [59]

    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

  52. [60]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11794– 11803, 2021. 2

    2021

  53. [61]

    You only learn one representation: Unified network for multiple tasks

    Chien-Yao Wang, I-Hau Yeh, and Hongpeng Liao. You only learn one representation: Unified network for multiple tasks. J. Inf. Sci. Eng., 39:691–709, 2021. 2

    2021

  54. [62]

    Yolov9: Learning what you want to learn using programmable gradi- ent information.ArXiv, abs/2402.13616, 2024

    Chien-Yao Wang, I-Hau Yeh, and Hongpeng Liao. Yolov9: Learning what you want to learn using programmable gradi- ent information.ArXiv, abs/2402.13616, 2024. 2

    2024 arXiv

  55. [63]

    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

    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. 2, 3

    2023

  56. [64]

    Second: Sparsely embed- ded convolutional detection.Sensors (Basel, Switzerland), 18, 2018

    Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embed- ded convolutional detection.Sensors (Basel, Switzerland), 18, 2018. 2, 9

    2018

  57. [65]

    Pixor: Real- time 3d object detection from point clouds.2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7652–7660, 2018

    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

    2018

  58. [66]

    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. 2

    2023

  59. [67]

    3dssd: Point-based 3d single stage object detector.2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11037–11045, 2020

    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

  60. [68]

    Center- based 3d object detection and tracking.CVPR, 2021

    Tianwei Yin, Xingyi Zhou, and Philipp Kr ¨ahenb¨uhl. Center- based 3d object detection and tracking.CVPR, 2021. 2, 3, 9

    2021

  61. [69]

    Beverse: Uni- fied perception and prediction in birds-eye-view for vision- centric autonomous driving.ArXiv, abs/2205.09743, 2022

    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. 2

    2022 arXiv

  62. [70]

    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

    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 14 A WML: An Open-Source ML-based Robotics Perception Framework to Deploy for ROS-ba...

    2018

  63. [71]

    autoware ml/

    Appendix 6.1. Contribution • Tanaka: Design whole architecture of AWML. • Tanaka, Samrat, Kok, Amadeusz, Kenzo, Minoda, Tomie, Uetake: Main contributor for AWML. • Zhang, Minoda: Contributor of constructing T4dataset. • Minoda, Yamashita, Horibe: Management of research and dev...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.