Pith. sign in

REVIEW 4 major objections 6 minor 52 references

Reflective Teacher: Semi-Supervised Multimodal 3D Object Detection in Bird's-Eye-View via Uncertainty Measure

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Reflective Teacher matches full-supervision 3D detection using only 25% of nuScenes labels and 22% of Waymo labels.

desk verdict Strong numbers, but the regularizer that makes this 'Reflective Teacher' is never isolated, and the forgetting metric can't carry the anti-forgetting claim. read the letter →

arxiv 2412.04337 v1 pith:H2E5QSTL submitted 2024-12-05 cs.CV

Saheli Hazra , Sudip Das , Rohit Choudhary , Arindam Das , Ganesh Sistu , Ciaran Eising , Ujjwal Bhattacharya This is my paper · ORCID
classification cs.CV
keywords semi-supervised3Dobjectdetectionbird's-eye-viewteacher-studentcatastrophicforgettinguncertaintyweightingLiDAR-camerafusionpseudo-labelingautonomousdriving
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 claims that semi-supervised training can eliminate most of the annotation cost of 3D object detection for autonomous driving: its Reflective Teacher pipeline, trained on only 25% of the labeled nuScenes data and 22% of the labeled Waymo data, reaches the same bird's-eye-view detection accuracy as a model trained on the full labeled set. The authors identify catastrophic forgetting in the teacher network, caused by exponential-moving-average weight updates, as the main obstacle to pseudo-label quality, and they propose a reflective regularizer plus an uncertainty-weighted loss to remove that obstacle. They also introduce a geometry-aware fusion module that aligns camera and LiDAR bird's-eye-view feature statistics before detection. If the claim holds, the method would cut the cost of 3D bounding-box annotation by roughly a factor of four on these benchmarks.

What carries the argument

The core object is the reflective regularizer in Equation (9), which combines an L1 consistency loss between the exponential-moving-average-updated teacher and the previous teacher on unlabeled data with a quadratic penalty on parameter changes, weighted by importance scores. Those scores are computed as the gradient of the previous teacher's output norm with respect to each parameter, so parameters whose perturbation would most change the teacher's predictions are the ones most strongly pinned down. The second load-bearing mechanism is the uncertainty measure in Equation (11), which assigns each region proposal a noise score based on the classification score and its maximum IoU with a pseudo-label, then multiplies the unsupervised regression loss by (1 minus the noise score) so that uncertain proposals contribute less. Together these mechanisms are meant to keep the teacher's knowledge stable while letting the student learn from both labeled and pseudo-labeled data.

What would settle it

Train the proposed model identically with and without the reflective regularizer, holding all other components fixed, and compare both mAP and the forgetting score at matched recall levels. If the regularizer is responsible for the paper's results, the version without it should show substantially higher forgetting when overall accuracy is held equal, and the version with it should retain more of the supervised model's correct detections; if the forgetting scores are the same after matching accuracy, the anti-forgetting claim is not supported.

Watch

Extended reading notes

Core claim

Section 4.1 states the central result directly: the model trained on only 25% labeled data achieves an equivalent result to the fully labeled dataset. On the nuScenes test set the semi-supervised Reflective Teacher reaches 79.46% mean average precision (mAP) against 79.3% for the same architecture trained on 100% of labels, and on Waymo L1 it reaches 85.31% mAP with 22% labels, comparable to the fully supervised model. The paper attributes this to three components: a reflective regularizer that protects parameters important to the previous teacher's predictions during the exponential-moving-average update, an uncertainty measure that down-weights region proposals whose assigned pseudo-labels are likely wrong, and GA-BEVFusion, which aligns camera and LiDAR feature distributions in bird's-eye-view space before fusing them. The same components also lift fully supervised performance above prior state-of-the-art on both datasets.

Load-bearing premise

The anti-forgetting evidence rests on a forgetting metric that is not normalized by overall detection accuracy, so the low forgetting scores could occur even if the reflective regularizer itself contributes nothing; the load-bearing assumption is that this metric measures what the regularizer is claimed to do.

Editorial extensions

If this is right

  • With 25% of nuScenes labels, the model reaches 79.46% mAP on the test set, statistically equivalent to the 79.3% mAP of the fully supervised model, and with 22% of Waymo labels it reaches 85.31% mAP at L1 difficulty.
  • The catastrophic forgetting score, denoted Δ, drops to 0.4–0.9% for Reflective Teacher across label fractions, versus roughly 7–10% for the Unbiased Teacher and Active Teacher baselines.
  • Ablation results show each component contributes: the uncertainty measure improves mAP at every label percentage (for example, from 33.82 to 47.23 at 5%), GA-BEVFusion adds 2.29% mAP over naive feature concatenation in the fully supervised setting, and perspective supervision adds 4–5% mAP.
  • The fully supervised version also surpasses prior state-of-the-art multimodal detectors on the nuScenes test set (79.3% mAP, 80.4% NDS) and the Waymo test set (85.9% mAP and 84.1% mAPH at L1).
  • The fusion module increases the IoU between camera-only and LiDAR-only bird's-eye-view regressor outputs to 61.23% on nuScenes and 67.74% on Waymo, indicating improved cross-modal alignment.

Reading between the lines

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

  • The 25% and 22% equivalence claims are measured against the authors' own fully supervised model, not against every published method; the practical annotation saving could be larger or smaller depending on the chosen labeled subset and the baseline, so a randomized multi-seed study would clarify how robust the factor-of-four claim is.
  • Because the forgetting metric computes the fraction of the supervised model's correct detections that survive after semi-supervised training, and the semi-supervised model has much higher overall accuracy, a low Δ may be partly a side effect of higher mAP rather than direct evidence of memory preservation; a matched-recall comparison would separate these effects.
  • The same reflective regularizer could be applied to other exponential-moving-average-based semi-supervised detectors, both 2D and 3D, and the same feature-statistics alignment could be tested on other modality pairs, which are testable extensions the paper does not explore.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes a semi-supervised teacher–student framework for multi-modal BEV 3D object detection. The contributions are a 'Reflective Teacher' regularizer intended to prevent catastrophic forgetting during EMA-based teacher updates, an uncertainty-weighted RPN loss for filtering noisy pseudo-labels, and a Geometry-Aware BEV Fusion module for aligning camera and LiDAR BEV features. Experiments on nuScenes and Waymo report state-of-the-art fully supervised results and claim that training on only 25% (nuScenes) or 22% (Waymo) of labeled data matches fully supervised performance, with very low measured forgetting.

Significance. If the central claim holds, the method would reduce 3D detection annotation cost by roughly 4x, which is practically important for autonomous driving. The paper has useful ingredients: it evaluates on two large datasets, abalates perspective supervision and the uncertainty term, and compares against Unbiased Teacher and Active Teacher on the same architecture. However, the core novelty — the reflective regularizer in Eq. (9) — is never isolated in any experiment, and the forgetting metric used to support the anti-forgetting claim is confounded by overall detection accuracy. The reported numbers come from a single unreleased implementation with no hyperparameters or training details, so the load-bearing claims are currently not independently verifiable. With the missing ablation and reporting fixed, the paper could make a solid contribution; in its present form the central mechanism is unevidenced.

major comments (4)
  1. [§4.2, Eq. (9), Tables 5–6] No experiment disables the reflective regularizer. Table 5 varies perspective supervision and the uncertainty term, and Table 6 varies only the uncertainty term; neither compares η>0 with η=0 while holding the active-sampling, EMA, and uncertainty components fixed. The comparison against Unbiased Teacher and Active Teacher in Table 3 changes multiple components at once. Therefore the 25%-labeled result (79.46 mAP, Table 3) is not shown to depend on the paper's central novelty. An ablation that sets η=0 in Eq. (9) while keeping the rest of the pipeline unchanged is required before the equivalence claim can be attributed to Reflective Teacher.
  2. [§4.1, Table 3] The forgetting metric Δ = (V1 − |V1∩V2|)/V1 is not adjusted for detection accuracy. Because the compared semi-supervised models reach very different mAP levels (e.g., 79.46 vs. 55.94 for Reflective Teacher and Unbiased Teacher at 25% nuScenes), a larger V1∩V2 can result simply from higher recall, not from the regularizer preserving prior knowledge. The paper should control for the detection operating point (e.g., a fixed number of top-k detections per class or matched recall) or report V1 and V2 separately per class. As written, the low forgetting percentages do not establish that the reflective regularizer is responsible for knowledge retention.
  3. [§3.3, Eqs. (8)–(9)] The teacher update procedure is ambiguous. The text states that teacher parameters are first updated by EMA and then 'updated iteratively through backpropagation of L_ref', but L_ref depends on M_TEMA_k and the previous teacher network. It is not specified which network's parameters receive gradients from L_ref, whether the student is updated on L_stud concurrently, or how the EWC-style importance weights Φ are accumulated across iterations. This is a central mechanism of the paper and must be specified precisely, ideally with an algorithm box or pseudo-code, before the method can be reproduced or evaluated.
  4. [§4, Tables 1–6] The experimental setup is under-reported. No values are given for the EMA coefficient α, regularizer weight η, IoU threshold Δ, uncertainty sigmoid parameter β, loss weights λ, γ, κ, the active-sampling top-m size, or the number of training iterations. No data split details, number of random seeds, or standard deviations are provided. Since the central equivalence claim (79.46 vs. 79.3 mAP at 25% labeled data) depends on a difference comparable to run-to-run variation, these details — ideally with multiple runs — are needed to support the headline claim.
minor comments (6)
  1. [§4.2, Table 5] The text after Table 5 reports an 'incremental improvement of 0.2% in mAP' for adding uncertainty, while Table 5 shows 76.1→79.46 mAP (a 3.36-point difference); please correct the text/table inconsistency or clarify that 0.2% refers to the difference from the fully supervised result.
  2. [§4.2, final paragraph] The statement that BEVFusion and LIFT 'achieved equivalent accuracy with more than 35% labeled data' is not supported by any table or training details; it should be either reported in full or removed.
  3. [§3.4, Eq. (11)] The sigmoid parameter β is introduced but never defined or given a value; the notation β' is also confusing because β' is defined in terms of β rather than being an independent exponent.
  4. [§4.1] The Waymo paragraph contains the typo 'Waymo attends its equivalent supervised mAP value' — it should read 'attains'; Table 5 also has a formatting typo '82. 65'.
  5. [Figure 1] The caption contains an incomplete sentence ('Active for and inactive during backward pass for'); please rephrase to a complete sentence.
  6. [Abstract and §4.1] The phrase 'equivalent performance with only 25%' should explicitly state that the comparison is to the fully supervised model with the same architecture and that no statistical significance is reported; this would make the claim more precise.

Circularity Check

1 steps flagged · score 2.0 of 10

No circular derivation: the 25%-label equivalence is an empirical benchmark result; only the forgetting metric Δ is partially by construction because the regularizer directly optimizes prediction consistency.

  1. other [Section 4.1 (forgetting metric definition) and Section 3.3, Eq. (9)]
    "To address potential discrepancies between pseudo-labels generated by the M TEM A k and those produced by the previous teacher network M T k−1 when unlabeled data is processed, a regularizer is incorporated into the loss function, as depicted in Eq. (9), to penalize the changes in the parameters. ..."

    The regularizer in Eq. (9) explicitly minimizes the L1 distance between the current EMA teacher's outputs and the previous teacher's outputs on unlabeled data, i.e., it optimizes prediction consistency across teacher updates. The forgetting metric Δ is then defined as the fraction of the supervised model's correct detections lost by the semi-supervised model. Because the semi-supervised model is initialized from the supervised model and the regularizer directly penalizes drift from the previous teacher, a small Δ is substantially the optimized objective itself, not an independent measure of knowledge retention. The metric is additionally not adjusted for the large mAP increase, which mechanically enlarges |V1∩V2| through recall.

full rationale

The paper's central claims—equivalent mAP/NDS with 25%/22% labeled data—are empirical benchmark results on held-out nuScenes and Waymo test sets (Tables 1-3, 5-6), not derivations from the method's definitions. The reflective regularizer (Eq. 9), uncertainty reweighting (Eq. 12), and GA-BEVFusion (Eqs. 2-3) are implemented from external, cited mechanisms (MAS [17], Active Teacher [19], BEVFormer v2 [11]) and are compared against published SOTA and internal ablations. There is no fitted parameter renamed as a prediction, no load-bearing self-citation, and no imported uniqueness theorem. The regularizer itself is not independently ablated, so the headline gain cannot be fully attributed to it; this is a completeness issue, not circularity. The only mild circularity is in the forgetting evaluation: Δ is a detection-level proxy of the very prediction-consistency objective that the regularizer optimizes, and it is mAP-confounded. This affects the strength of the anti-forgetting evidence, not the validity of the mAP results; it is an evaluation-alignment issue rather than a circular derivation. Accordingly, score 2.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

No new physical entities are proposed; the contributions are training and fusion techniques. The free parameters listed are those most directly tied to the method's behavior and claimed benefits, and none of them is reported in the paper.

free parameters (7)
  • EMA coefficient α = not reported
    Controls teacher update speed; the paper describes EMA but does not state α, and its choice affects pseudo-label stability.
  • Regularizer weight η = not reported
    Scales the memory-aware synapses penalty in Eq. (9); no value or schedule given, yet it controls the strength of the anti-forgetting mechanism.
  • IoU threshold Δ = not reported
    Determines whether a region proposal is positive or negative for uncertainty weighting in Eq. (11); threshold not specified.
  • Uncertainty sigmoid parameter β = not reported
    Appears in β' in Eq. (11) but β is never defined or given a value.
  • Loss weights λ, γ, κ = not reported
    Balance perspective loss, alignment loss, and unsupervised loss in Eqs. (1),(4),(6),(7); no values stated.
  • Active sampling top-m size = not reported
    The number of pseudo-labeled samples added per iteration from Active Teacher sampling is not specified.
  • VGG alignment layers = relu1_1, relu2_1, relu3_1, relu4_1
    The paper states these layers are used in the alignment loss; this is a design choice affecting the fusion loss.
assumptions (5)
  • domain assumption EMA teacher updates cause catastrophic forgetting in SSOD
    The paper motivates the reflective regularizer by this claim (Section 1, Section 3.3), but it is not empirically demonstrated in the paper itself; EMA is usually used to reduce drift.
  • domain assumption Gradient magnitude on unlabeled inputs measures parameter importance for preserving previous knowledge
    Borrowed from Memory Aware Synapses [17] and used in Eq. (8); assumes this importance estimate transfers to the teacher-student setting.
  • domain assumption Aligning mean and variance of camera BEV features to LiDAR BEV features reduces geometric misalignment
    The GA-BEVFusion design (Eq. 2) assumes first-order statistics alignment is sufficient to handle spatial misalignment, which is not geometrically grounded.
  • domain assumption VGG-19 features of BEV feature maps can supervise alignment via mean/variance matching
    The alignment loss (Eq. 3) applies a VGG network to multi-channel BEV features without describing how the channel dimension is reduced to match VGG input; assumes this is a valid perceptual alignment.
  • domain assumption Active Teacher's sampling score β ranks the most useful unlabeled samples
    The paper adopts the active sampling strategy from [19] without re-validation in this setting.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reflective Teacher: Semi-Supervised Multimodal 3D Object Detection in Bird's-Eye-View via Uncertainty Measure." pith.science (2026). https://pith.science/paper/H2E5QSTL

@misc{pith2026241204337,
  author       = {Pith},
  title        = {Pith review of: Reflective Teacher: Semi-Supervised Multimodal 3D Object Detection in Bird's-Eye-View via Uncertainty Measure},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H2E5QSTL}},
  note         = {Machine review of arXiv:2412.04337}
}
read the original abstract

Applying pseudo labeling techniques has been found to be advantageous in semi-supervised 3D object detection (SSOD) in Bird's-Eye-View (BEV) for autonomous driving, particularly where labeled data is limited. In the literature, Exponential Moving Average (EMA) has been used for adjustments of the weights of teacher network by the student network. However, the same induces catastrophic forgetting in the teacher network. In this work, we address this issue by introducing a novel concept of Reflective Teacher where the student is trained by both labeled and pseudo labeled data while its knowledge is progressively passed to the teacher through a regularizer to ensure retention of previous knowledge. Additionally, we propose Geometry Aware BEV Fusion (GA-BEVFusion) for efficient alignment of multi-modal BEV features, thus reducing the disparity between the modalities - camera and LiDAR. This helps to map the precise geometric information embedded among LiDAR points reliably with the spatial priors for extraction of semantic information from camera images. Our experiments on the nuScenes and Waymo datasets demonstrate: 1) improved performance over state-of-the-art methods in both fully supervised and semi-supervised settings; 2) Reflective Teacher achieves equivalent performance with only 25% and 22% of labeled data for nuScenes and Waymo datasets respectively, in contrast to other fully supervised methods that utilize the full labeled dataset.

Figures

Figures reproduced from arXiv: 2412.04337 by the authors.

Figure 1
Figure 1. The architecture of our proposed end-to-end multi-modal framework for 3D object detection in BEV. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Data flow sequence and interaction of teacher-student [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative visual results of our 3D object detection model on nuScenes [ [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Non-overlapping bounding boxes due to feature mis [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 45 canonical work pages

  1. [1]

    Deep learning-based motion pre- diction leveraging autonomous driving datasets: State- of-the-art,

    F. A. Barrios et al., “Deep learning-based motion pre- diction leveraging autonomous driving datasets: State- of-the-art,” IEEE Access, 2024. 1

    work page 2024

  2. [2]

    Path planning techniques for au- tonomous vehicles,

    Z. Wu et al. , “Path planning techniques for au- tonomous vehicles,” in AIP Conference Proceedings, vol. 3144, AIP Publishing, 2024. 1

    work page 2024

  3. [3]

    Bevfusion: Multi-task multi-sensor fu- sion with unified bird’s-eye view representation,

    Z. Liu et al., “Bevfusion: Multi-task multi-sensor fu- sion with unified bird’s-eye view representation,” in IEEE ICRA, pp. 2774–2781, 2023. 1, 2, 6, 8

    work page 2023

  4. [4]

    Ea-lss: Edge-aware lift-splat-shot framework for 3d bev object detection,

    H. Hu et al. , “Ea-lss: Edge-aware lift-splat-shot framework for 3d bev object detection,”arXiv preprint arXiv:2303.17895, 2023. 1, 2, 6

    arXiv 2023

  5. [5]

    Fisheye Camera and Ultrasonic Sensor Fusion For Near-Field Obstacle Perception in Bird's-Eye-View

    A. Das et al., “Fisheye camera and ultrasonic sensor fusion for near-field obstacle perception in bird’s-eye- view,”arXiv preprint arXiv:2402.00637, 2024. 1

    work page Pith review arXiv 2024

  6. [6]

    Spatio-contextual deep network- based multimodal pedestrian detection for au- tonomous driving,

    K. Dasgupta et al., “Spatio-contextual deep network- based multimodal pedestrian detection for au- tonomous driving,” IEEE TITS, 2022. 1

    work page 2022

  7. [7]

    Revisiting modality imbalance in mul- timodal pedestrian detection,

    A. Das et al., “Revisiting modality imbalance in mul- timodal pedestrian detection,” in 2023 IEEE ICIP , pp. 1755–1759, 2023. 1

    work page 2023

  8. [8]

    Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection,

    J. Yin et al., “Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection,” in Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pp. 14905–14915,

    work page

Show all 52 references
  1. [9]

    Graphbev: Towards robust bev feature alignment for multi-modal 3d object detection,

    Z. Song et al., “Graphbev: Towards robust bev feature alignment for multi-modal 3d object detection,” arXiv preprint arXiv:2403.11848, 2024. 1

    2024 arXiv

  2. [10]

    Graphalign++: An accurate feature alignment by graph matching for multi-modal 3d ob- ject detection,

    Z. Song et al. , “Graphalign++: An accurate feature alignment by graph matching for multi-modal 3d ob- ject detection,” IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1

    2023

  3. [11]

    Bevformer v2: Adapting mod- ern image backbones to bird’s-eye-view recognition via perspective supervision,

    C. Yang et al. , “Bevformer v2: Adapting mod- ern image backbones to bird’s-eye-view recognition via perspective supervision,” in IEEE/CVF CVPR , pp. 17830–17839, 2023. 1, 2, 6

    2023

  4. [12]

    nuscenes: A multimodal dataset for autonomous driving,

    H. Caesar et al. , “nuscenes: A multimodal dataset for autonomous driving,” in IEEE/CVF CVPR , pp. 11621–11631, 2020. 1, 2, 5, 6, 7

    2020

  5. [13]

    Scalability in perception for au- tonomous driving: Waymo open dataset,

    P. Sun et al. , “Scalability in perception for au- tonomous driving: Waymo open dataset,” in Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 2446–2454, 2020. 1, 2, 5, 6, 7

    2020

  6. [14]

    Unbiased teacher for semi- supervised object detection,

    Y .-C. Liu et al. , “Unbiased teacher for semi- supervised object detection,” arXiv preprint arXiv:2102.09480, 2021. 1, 7, 8

    2021 arXiv

  7. [15]

    Monocular 3d object detection with lidar guided semi supervised active learning,

    A. Hekimoglu et al., “Monocular 3d object detection with lidar guided semi supervised active learning,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pp. 2346–2355,

  8. [16]

    Alwod: active learning for weakly- supervised object detection,

    Y . Wang et al., “Alwod: active learning for weakly- supervised object detection,” in Proceedings of the IEEE/CVF international conference on computer vi- sion, pp. 6459–6469, 2023. 1

    2023

  9. [17]

    Memory aware synapses: Learning what (not) to forget,

    R. Aljundi et al., “Memory aware synapses: Learning what (not) to forget,” in ECCV, pp. 139–154, 2018. 1, 4, 5

    2018

  10. [18]

    Efficient lifelong learning with a- gem,

    A. Chaudhry et al., “Efficient lifelong learning with a- gem,” in International Conference on Learning Rep- resentations, 2018. 1

    2018

  11. [19]

    Active teacher for semi-supervised object detection,

    P. Mi et al., “Active teacher for semi-supervised object detection,” in IEEE/CVF CVPR , pp. 14482–14491,

  12. [20]

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

    Y . Zhou et al. , “V oxelnet: End-to-end learning for point cloud based 3d object detection,” inIEEE CVPR, pp. 4490–4499, 2018. 2, 7

    2018

  13. [21]

    Object as query: Lifting any 2d ob- ject detector to 3d detection,

    Z. Wang et al., “Object as query: Lifting any 2d ob- ject detector to 3d detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pp. 3791–3800, 2023. 2, 6

    2023

  14. [22]

    Cvfnet: Real-time 3d object detection by learning cross view features,

    J. Gu et al., “Cvfnet: Real-time 3d object detection by learning cross view features,” in2022 IEEE/RSJ IROS, pp. 568–574, IEEE, 2022. 2

    2022

  15. [23]

    Transfusion: Robust lidar-camera fu- sion for 3d object detection with transformers,

    X. Bai et al., “Transfusion: Robust lidar-camera fu- sion for 3d object detection with transformers,” in IEEE CVPR, pp. 1090–1099, 2022. 2, 6

    2022

  16. [24]

    Futr3d: A unified sensor fusion frame- work for 3d detection,

    X. Chen et al., “Futr3d: A unified sensor fusion frame- work for 3d detection,” in IEEE/CVF CVPR, pp. 172– 181, 2023. 2, 6

    2023

  17. [25]

    Bevfusion: A simple and robust lidar- camera fusion framework,

    T. Liang et al., “Bevfusion: A simple and robust lidar- camera fusion framework,” NIPS, vol. 35, pp. 10421– 10434, 2022. 2, 6

    2022

  18. [26]

    Lift: Learning 4d lidar image fusion transformer for 3d object detection,

    Y . Zeng et al., “Lift: Learning 4d lidar image fusion transformer for 3d object detection,” in IEEE/CVF CVPR, pp. 17172–17181, 2022. 2, 6, 8

    2022

  19. [27]

    Bevfusion4d: Learning lidar-camera fu- sion under bird’s-eye-view via cross-modality guid- ance and temporal aggregation,

    H. Cai et al., “Bevfusion4d: Learning lidar-camera fu- sion under bird’s-eye-view via cross-modality guid- ance and temporal aggregation,” arXiv preprint arXiv:2303.17099, 2023. 2, 6

    2023 arXiv

  20. [28]

    Objectfusion: Multi-modal 3d object detection with object-centric fusion,

    Q. Cai et al., “Objectfusion: Multi-modal 3d object detection with object-centric fusion,” in IEEE/CVF ICCV, pp. 18067–18076, 2023. 2, 6

    2023

  21. [29]

    Sess: Self-ensembling semi- supervised 3d object detection,

    N. Zhao et al. , “Sess: Self-ensembling semi- supervised 3d object detection,” in IEEE/CVF CVPR, pp. 11079–11087, 2020. 2

    2020

  22. [30]

    Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results,

    A. Tarvainen et al. , “Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results,” NIPS, vol. 30,

  23. [31]

    3dioumatch: Leveraging iou pre- diction for semi-supervised 3d object detection,

    H. Wang et al. , “3dioumatch: Leveraging iou pre- diction for semi-supervised 3d object detection,” in IEEE/CVF CVPR, pp. 14615–14624, 2021. 2

    2021

  24. [32]

    Semi-supervised 3d object detection with proficient teachers,

    J. Yin et al. , “Semi-supervised 3d object detection with proficient teachers,” in ECCV, pp. 727–743, Springer, 2022. 2

    2022

  25. [33]

    Combating noise: semi-supervised learning by region uncertainty quantification,

    Z. Wang et al. , “Combating noise: semi-supervised learning by region uncertainty quantification,” NIPS, vol. 34, pp. 9534–9545, 2021. 2

    2021

  26. [34]

    Reliable student: Addressing noise in semi-supervised 3d object detection,

    F. Nozarian et al. , “Reliable student: Addressing noise in semi-supervised 3d object detection,” inIEEE CVPR, pp. 4980–4989, 2023. 2

    2023

  27. [35]

    Internimage: Exploring large-scale vision foundation models with deformable convolu- tions,

    W. Wang et al., “Internimage: Exploring large-scale vision foundation models with deformable convolu- tions,” in IEEE CVPR, pp. 14408–14419, 2023. 2

    2023

  28. [36]

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

    T. Wang et al., “Fcos3d: Fully convolutional one-stage monocular 3d object detection,” in IEEE/CVF ICCV, pp. 913–922, 2021. 2

    2021

  29. [37]

    Deformable detr: Deformable transformers for end-to-end object detection,

    X. Zhu et al. , “Deformable detr: Deformable transformers for end-to-end object detection,” arXiv preprint arXiv:2010.04159, 2020. 2

    2010 arXiv

  30. [38]

    Very deep convolutional net- works for large-scale image recognition,

    K. Simonyan et al. , “Very deep convolutional net- works for large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014. 3

    2014 arXiv

  31. [39]

    V oxel r-cnn: Towards high perfor- mance voxel-based 3d object detection,

    J. Deng et al. , “V oxel r-cnn: Towards high perfor- mance voxel-based 3d object detection,” in AAAI, vol. 35, pp. 1201–1209, 2021. 4, 5

    2021

  32. [40]

    Center-based 3d object detection and tracking,

    T. Yin et al. , “Center-based 3d object detection and tracking,” in Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition , pp. 11784–11793, 2021. 6

    2021

  33. [41]

    Bev-io: Enhancing bird’s-eye-view 3d detection with instance occupancy,

    Z. Zhang et al., “Bev-io: Enhancing bird’s-eye-view 3d detection with instance occupancy,” arXiv preprint arXiv:2305.16829, 2023. 6

    2023 arXiv

  34. [42]

    Detr3d: 3d object detection from multi-view images via 3d-to-2d queries,

    Y . Wang et al. , “Detr3d: 3d object detection from multi-view images via 3d-to-2d queries,” in Confer- ence on Robot Learning , pp. 180–191, PMLR, 2022. 6

    2022

  35. [43]

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

    Z. Li et al. , “Bevformer: Learning bird’s-eye- view representation from multi-camera images via spatiotemporal transformers,” in ECCV, pp. 1–18, Springer, 2022. 6, 7

    2022

  36. [44]

    Fast-bev: Towards real-time on- vehicle bird’s-eye view perception,

    B. Huang et al. , “Fast-bev: Towards real-time on- vehicle bird’s-eye view perception,” arXiv preprint arXiv:2301.07870, 2023. 6

    2023 arXiv

  37. [45]

    Deepinteraction: 3d object detection via modality interaction,

    Z. Yang et al., “Deepinteraction: 3d object detection via modality interaction,” NIPS, vol. 35, pp. 1992– 2005, 2022. 6

    1992

  38. [46]

    Pointpillars: Fast encoders for object detection from point clouds,

    A. H. Lang et al. , “Pointpillars: Fast encoders for object detection from point clouds,” in IEEE/CVF CVPR, pp. 12697–12705, 2019. 6

    2019

  39. [47]

    Pillarnet: Real-time and high- performance pillar-based 3d object detection,

    G. Shi et al. , “Pillarnet: Real-time and high- performance pillar-based 3d object detection,” in Eu- ropean Conference on Computer Vision , pp. 35–52, Springer, 2022. 6

    2022

  40. [48]

    Super sparse 3d object detection,

    L. Fan et al., “Super sparse 3d object detection,”IEEE transactions on pattern analysis and machine intelli- gence, vol. 45, no. 10, pp. 12490–12505, 2023. 6

    2023

  41. [49]

    Centerformer: Center-based trans- former for 3d object detection,

    Z. Zhou et al. , “Centerformer: Center-based trans- former for 3d object detection,” in European Con- ference on Computer Vision , pp. 496–513, Springer,

  42. [50]

    Clusterformer: Cluster-based trans- former for 3d object detection in point clouds,

    Y . Pei et al. , “Clusterformer: Cluster-based trans- former for 3d object detection in point clouds,” inPro- ceedings of the IEEE/CVF International Conference on Computer Vision, pp. 6664–6673, 2023. 6

    2023

  43. [51]

    Pv-rcnn++: Point-voxel feature set ab- straction with local vector representation for 3d object detection,

    S. Shi et al., “Pv-rcnn++: Point-voxel feature set ab- straction with local vector representation for 3d object detection,” International Journal of Computer Vision, vol. 131, no. 2, pp. 531–551, 2023. 6

    2023

  44. [52]

    Afdetv2: Rethinking the necessity of the second stage for object detection from point clouds,

    Y . Huet al., “Afdetv2: Rethinking the necessity of the second stage for object detection from point clouds,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, pp. 969–979, 2022. 6

    2022

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.