Pith. sign in

REVIEW 4 major objections 4 minor 35 references

INTACT: Inducing Noise Tolerance through Adversarial Curriculum Training for LiDAR-based Safety-Critical Perception and Autonomy

T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read INTACT claims that meta-learned saliency maps, used to target adversarial curriculum perturbations, improve LiDAR perception robustness by up to 20 percentage points across detection, tracking, and classification.

desk verdict A plausible and consistently positive robustness recipe, but the headline numbers don't survive arithmetic and the novel mechanism isn't cleanly isolated. read the letter →

arxiv 2502.01896 v1 pith:UVYTC2F7 submitted 2025-02-04 cs.CV cs.RO

classification cs.CVcs.RO
keywords LiDARpointcloudsadversarialcurriculumtrainingmeta-learningsaliencymapsrobustnesstosensornoise3Dobjectdetectiontrackingcloudclassification
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

INTACT is a two-phase training framework that makes 3D LiDAR perception models tolerate incomplete and corrupted point clouds. In the first phase, a teacher network is meta-trained across noisy classification tasks so that its input gradients become saliency maps marking the points most critical to a prediction. In the second phase, a student network is trained against an adversarial curriculum that drops high-saliency points and injects Gaussian noise with rising intensity, while a gradient-alignment loss keeps the student's focus close to the teacher's. The paper reports that this combination lifts robustness beyond standard adversarial or curriculum training alone, recovering KITTI tracking MOTA from 64.1% to 75.1% under a 50% point drop and KITTI mAP from 49.3% to 70.9% under Gaussian noise. Such recovery matters because LiDAR degradations from weather, occlusion, or sensor faults otherwise cascade into perception failures.

What carries the argument

The load-bearing mechanism is the teacher's gradient saliency map, computed as $\nabla_x f_T(x)$ from the meta-trained teacher, coupled to a curriculum that drops the most salient points (starting at 90% of points and decreasing to 50%) while the Gaussian noise variance $\sigma_t$ grows. The student is trained with a four-term objective: cross-entropy on clean inputs, a robustness loss that a discriminator maximizes, the alignment loss $\mathcal{L}_{\mathrm{diff}} = \|\nabla_x f_S(x) - \nabla_x f_T(x)\|_2$ that transfers the teacher's focus, and curriculum-scheduled versions of the robustness and alignment terms. This couples where to perturb (saliency) with when to perturb (curriculum), so the student is pushed hardest on the points the teacher identifies as decisive.

What would settle it

Run the same adversarial curriculum on the same student and data, but rank points for dropping and corrupting by a random or uniform rule instead of by the teacher's saliency; if tracking MOTA under the 50% point drop still reaches roughly 75.1% on KITTI, the saliency-transfer mechanism is not what carries the improvement.

Watch

Extended reading notes

Core claim

The paper's central claim is that the combination of meta-learned saliency guidance and adversarial curriculum training, not either ingredient alone, produces the robustness gain. After meta-training, the teacher's gradient map tells the perturbation generator exactly which points to drop or corrupt; the curriculum first perturbs up to 90% of the most salient points with mild noise and then, as training proceeds, perturbs fewer points with stronger noise. Across detectors PointPillar, SECOND, CenterFormer, and TED on KITTI and Argoverse, INTACT consistently beats ACT-only, and on ModelNet40 it restores Point Transformer to 92.7% accuracy under 50% point drop from a degraded 87.2%, close to its 93.8% clean baseline. The authors read these results as evidence that targeted perturbation of critical regions, enforced through the gradient-alignment loss of Eq. 5, teaches representations that survive structured noise without any extra cost at inference.

Load-bearing premise

Everything rests on the teacher's map of which input points matter most transferring its meaning to a different student architecture and task; if those maps do not point at the student's genuinely critical points, INTACT is just a curriculum over random dropout and noise.

Editorial extensions

If this is right

  • Detectors and trackers trained with INTACT retain most of their clean accuracy under a 50% point drop or Gaussian noise, with no change to the deployed architecture.
  • The same schedule improves all four detection backbones, all five classification backbones, and both driving datasets, so the paper presents it as a task- and architecture-agnostic recipe.
  • On Argoverse, bus detection mean average precision under Gaussian noise rises from 17.4% to 32.3%, showing large absolute gains on the class with the weakest noisy baseline.
  • Because the teacher and discriminator are used only during training, the deployed student model carries no extra inference cost.

Reading between the lines

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

  • If the teacher's saliency maps transfer as claimed, the meta-learning phase could be run once per sensor configuration and reused to harden any downstream student, turning INTACT into a plug-in training schedule rather than a new model family.
  • The evaluations use synthetic degradations (uniform point drop and Gaussian noise), so the natural next test is the same curriculum against physically recorded rain, fog, or sensor-misalignment corruptions; nothing in the paper rules out that the gains shrink or grow there.
  • A useful diagnostic would be to track the gradient-alignment loss $\mathcal{L}_{\mathrm{diff}}$ across training: if it correlates with downstream robustness, it could serve as a stopping criterion or a cheap proxy for noise tolerance during development.
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 / 4 minor

Summary. The manuscript presents INTACT, a two-phase framework for improving the robustness of LiDAR-based 3D perception models. In Phase I, a teacher network is meta-trained on classification tasks to produce input-gradient saliency maps. In Phase II, a student network is trained with an adversarial curriculum in which perturbations (point dropping and Gaussian noise) are guided by the teacher's saliency maps and a gradient-alignment loss. The authors evaluate INTACT on KITTI and Argoverse object detection, ModelNet40 classification, and YOLOv8n tracking, reporting consistent improvements over a baseline and over adversarial curriculum training (ACT) alone. The central empirical claim is that INTACT provides large robustness gains under 50% point drop and Gaussian noise, e.g., KITTI tracking mAP rising from 49.3% to 70.9% under Gaussian noise.

Significance. If the reported results are reproducible, the proposed combination of meta-learned saliency guidance with adversarial curriculum training would be a useful contribution to robust 3D perception. The evaluation spans multiple architectures, datasets, and tasks, which is commendable. However, the manuscript does not yet support its central attribution: no ablation isolates the saliency-guided perturbation mechanism from the curriculum schedule, and no standard adversarial-training baseline is included despite the abstract's claim to outperform such methods. The absence of error bars and the presence of numeric inconsistencies in the abstract further limit the present claims. The central idea is plausible and worth testing, but the evidence as presented is incomplete.

major comments (4)
  1. [Section III-B, Eq. (5); Tables I-IV] The core novelty of INTACT is the teacher saliency-guided perturbation mechanism, yet no ablation isolates it. Every INTACT result includes both the saliency-based point selection and the gradient-alignment loss L_diff from Eq. (5), so the reported gains cannot be attributed to the teacher's saliency maps. The comparison against ACT alone only supports the curriculum schedule. A control experiment with uniformly random point selection and no Eq. (5) is needed to validate the mechanism that distinguishes INTACT from plain ACT.
  2. [Abstract and Table IV] The abstract contains numeric claims that are inconsistent with the tables. It states an improvement of 'up to 20%' but Table IV shows KITTI Gaussian-noise mAP rising from 49.3% to 70.9%, a 21.6-point (43.8% relative) increase. The abstract also reports MOTA gains of 9.6% (64.1% to 75.1%) and 12.4% (52.5% to 73.7%), but the table deltas are 11.0 and 21.2 points, respectively. These numbers must be reconciled or corrected before the headline claims can be accepted.
  3. [Section IV, Tables I-IV] The abstract claims INTACT 'outperforms standard adversarial and curriculum training methods,' but the experiments only compare against a baseline and ACT; no standard adversarial training (e.g., a PGD-based or random-noise adversarial training baseline) is reported. Without that baseline, the claim of outperforming standard adversarial training is unsupported. Additionally, all results are reported as single numbers with no error bars or repeated-run statistics, so the reader cannot assess the variability or statistical significance of the reported improvements.
  4. [Section III-B, Eq. (5)] The gradient-alignment loss L_diff = ||∇_x f_S(x) − ∇_x f_T (x)||_2 is under-specified for the detection and tracking experiments. The teacher f_T is a meta-trained classifier, while the student f_S is a detection or tracking network (e.g., PointPillars, SECOND, YOLOv8n). The paper does not state how ∇_x f_S(x) is computed for these outputs, which output head is used, or how the dimensions of the teacher and student gradients are matched. If the loss cannot be computed as written for these architectures, the reported results are not reproducible.
minor comments (4)
  1. [Section II-A] The equations for P_total, P_laser, E_pulse, and P_ADC are unnumbered; numbering them would make the paper easier to reference.
  2. [Section IV-A] The text states that on Argoverse 'SECOND's Bus class mAP increases from 38.6% (baseline with noise) to 69.8% with INTACT,' but Table II shows the Noise + INTACT value for SECOND Bus is 38.6%, and the baseline with noise is 23.4%. This appears to be a typo and should be corrected.
  3. [Section IV-A and Table I] The architecture is referred to as 'PointPillar' in the text and Table I, but the reference [30] is titled 'PointPillars'; the name should be consistent.
  4. [Section III-B(iii)] The curriculum schedule (perturbed fraction decreasing from 90% to 50% and σ_t increasing by Δσ) is hand-specified without any sensitivity analysis or justification. Reporting how these values were selected, or ablating them, would strengthen the reproducibility of the method.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: INTACT's robustness gains are empirical comparisons on external benchmarks; the saliency-guided curriculum contains no fitted quantity that is then relabeled as a prediction.

full rationale

I walked the paper's derivation chain from Phase I meta-learning (Eq. 1) through teacher saliency-map generation (Eq. 2) to the student objective (Eqs. 3-5) and the curriculum perturbation schedule. Nothing in this chain defines a target result in terms of itself: the teacher's input gradients are computed from the teacher's own classifier, the student's gradient-alignment loss L_diff is a training regularizer rather than a predictive claim, and the reported MOTA/mAP improvements are measured against held-out KITTI, Argoverse, and ModelNet40 conditions. The perturbation families used during training (50% point drop, Gaussian noise with sigma = 0.1) are the same families used during evaluation, so the results demonstrate robustness to the trained noise types rather than to genuinely unseen noise; that is a generalization or external-validity limitation, not a definitional equivalence or fitted-input relabeling. The paper's self-citations ([16]-[18], [23]-[26]) appear in background and related-work contexts and are not load-bearing for the central mechanism; none is invoked as a uniqueness theorem or as the sole justification for the method's core premise. No parameter is fitted to a subset of data and then called a prediction, and no equation reduces to its own input by construction. The central empirical comparison is self-contained against external benchmarks, so no significant circularity is present.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

All components of the pipeline are defined by high-level equations; the concrete numbers that make the method work (loss weights, noise schedule, stage length, teacher meta-learning settings) are left unspecified, so the central claim rests on unstated implementation choices. The strongest domain assumption is that synthetic Gaussian noise and random point dropping, used in both training and evaluation, represent the LiDAR degradation modes the paper motivates. No new physical entities are introduced.

free parameters (6)
  • beta (loss weight for L_Robust and L_curr_Robust)
    Weights in Eq. 3 are introduced without values or sensitivity analysis; the balance between clean accuracy and robustness directly determines reported gains.
  • gamma (loss weight for L_diff and L_curr_diff)
    Gradient-alignment strength in Eq. 3 and Eq. 5 is hand-chosen; no ablation shows how teacher transfer depends on it.
  • Perturbed-fraction curriculum (90% down to 50%)
    Section III-B(iii) fixes the fraction of saliency-ranked points perturbed; no sweep or justification is given.
  • Gaussian noise schedule (sigma_t, Delta_sigma)
    Section III-B(iii) specifies sigma_{t+1} = sigma_t + Delta_sigma but does not give start, increment, or stage length; these control the difficulty ramp.
  • Curriculum stage count and iteration schedule
    Closed-loop training iterations and epoch counts per stage are unspecified, so the severity ramp is not reproducible.
  • Teacher meta-learning hyperparameters (task batch K, inner-loop steps, learning rates)
    Phase I is defined only by Eq. 1; no MAML or Reptile update details, task construction, or optimizer settings are given.
assumptions (3)
  • domain assumption Input gradients d f_T / d x (Eq. 2) reliably identify the points most critical to prediction across architectures and noise levels.
    The teacher's saliency maps are the sole guidance for perturbations; the paper provides no validation that gradient magnitude correlates with task-critical points for detection or tracking. Eq. 5 only forces alignment between teacher and student gradients, not correctness of the saliency.
  • domain assumption Gaussian noise with sigma = 0.1 and random 50% point dropping are representative of the LiDAR data-loss and data-corruption regimes described in Section II.
    Both training (Section III-B) and evaluation (Section IV) use the same synthetic perturbation families, so gains may reflect matching training and test noise rather than tolerance to realistic rain, fog, or multipath distortions.
  • standard math Gradient-based optimization of the objectives in Eqs. 1-5 is well-posed with the chosen losses and differentiable approximations.
    The paper assumes standard backpropagation and min-max training are valid for these point-cloud architectures, and that the curriculum terms are differentiable as written.

how reviews work

0 comments
Cite this review

Pith. "Pith review of INTACT: Inducing Noise Tolerance through Adversarial Curriculum Training for LiDAR-based Safety-Critical Perception and Autonomy." pith.science (2026). https://pith.science/paper/UVYTC2F7

@misc{pith2026250201896,
  author       = {Pith},
  title        = {Pith review of: INTACT: Inducing Noise Tolerance through Adversarial Curriculum Training for LiDAR-based Safety-Critical Perception and Autonomy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UVYTC2F7}},
  note         = {Machine review of arXiv:2502.01896}
}
read the original abstract

In this work, we present INTACT, a novel two-phase framework designed to enhance the robustness of deep neural networks (DNNs) against noisy LiDAR data in safety-critical perception tasks. INTACT combines meta-learning with adversarial curriculum training (ACT) to systematically address challenges posed by data corruption and sparsity in 3D point clouds. The meta-learning phase equips a teacher network with task-agnostic priors, enabling it to generate robust saliency maps that identify critical data regions. The ACT phase leverages these saliency maps to progressively expose a student network to increasingly complex noise patterns, ensuring targeted perturbation and improved noise resilience. INTACT's effectiveness is demonstrated through comprehensive evaluations on object detection, tracking, and classification benchmarks using diverse datasets, including KITTI, Argoverse, and ModelNet40. Results indicate that INTACT improves model robustness by up to 20% across all tasks, outperforming standard adversarial and curriculum training methods. This framework not only addresses the limitations of conventional training strategies but also offers a scalable and efficient solution for real-world deployment in resource-constrained safety-critical systems. INTACT's principled integration of meta-learning and adversarial training establishes a new paradigm for noise-tolerant 3D perception in safety-critical applications. INTACT improved KITTI Multiple Object Tracking Accuracy (MOTA) by 9.6% (64.1% -> 75.1%) and by 12.4% under Gaussian noise (52.5% -> 73.7%). Similarly, KITTI mean Average Precision (mAP) rose from 59.8% to 69.8% (50% point drop) and 49.3% to 70.9% (Gaussian noise), highlighting the framework's ability to enhance deep learning model resilience in safety-critical object tracking scenarios.

Figures

Figures reproduced from arXiv: 2502.01896 by the authors.

Figure 1
Figure 1. INTACT Overview: This figure shows overview of our proposed [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 15 canonical work pages

  1. [1]

    Robust localization of mobile robots considering reliability of lidar measurements,

    J. Kim and W. Chung, “Robust localization of mobile robots considering reliability of lidar measurements,” in 2018 IEEE International Confer- ence on Robotics and Automation (ICRA). IEEE, 2018, pp. 6491–6496

  2. [2]

    Extended kalman filter (ekf) design for vehicle position tracking using reliability function of radar and lidar,

    T. Kim and T.-H. Park, “Extended kalman filter (ekf) design for vehicle position tracking using reliability function of radar and lidar,” Sensors, vol. 20, no. 15, p. 4126, 2020

  3. [3]

    Uav for 3d mapping applications: a review,

    F. Nex and F. Remondino, “Uav for 3d mapping applications: a review,” Applied geomatics, vol. 6, pp. 1–15, 2014

  4. [4]

    L3-net: Towards learning based lidar localization for autonomous driving,

    W. Lu, Y . Zhou, G. Wan, S. Hou, and S. Song, “L3-net: Towards learning based lidar localization for autonomous driving,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2019, pp. 6389–6398

  5. [5]

    Lidar denoising methods in adverse environments: A review,

    J.-i. Park, S. Jo, H.-T. Seo, and J. Park, “Lidar denoising methods in adverse environments: A review,” IEEE Sensors Journal , 2025

  6. [6]

    Deep learning for lidar point clouds in autonomous driving: A review,

    Y . Li, L. Ma, Z. Zhong, F. Liu, M. A. Chapman, D. Cao, and J. Li, “Deep learning for lidar point clouds in autonomous driving: A review,” IEEE Transactions on Neural Networks and Learning Systems , vol. 32, no. 8, pp. 3412–3432, 2020

  7. [7]

    Towards deep learning models resistant to adversarial attacks,

    A. Madry, A. Makelov, L. Schmidt, D. Tsipras, and A. Vladu, “Towards deep learning models resistant to adversarial attacks,” arXiv preprint arXiv:1706.06083, 2017

  8. [8]

    Curriculum learning,

    Y . Bengio, J. Louradour, R. Collobert, and J. Weston, “Curriculum learning,” in Proceedings of the 26th annual international conference on machine learning , 2009, pp. 41–48

Show all 35 references
  1. [9]

    Model-agnostic meta-learning for fast adaptation of deep networks,

    C. Finn, P. Abbeel, and S. Levine, “Model-agnostic meta-learning for fast adaptation of deep networks,” in International conference on machine learning. PMLR, 2017, pp. 1126–1135

  2. [10]

    A two-stage clustering based 3d visual saliency model for dynamic scenarios,

    Y . Yang, B. Li, P. Li, and Q. Liu, “A two-stage clustering based 3d visual saliency model for dynamic scenarios,” IEEE Transactions on Multimedia, vol. 21, no. 4, pp. 809–820, 2018

  3. [11]

    Curriculum adversarial training,

    Q.-Z. Cai, M. Du, C. Liu, and D. Song, “Curriculum adversarial training,” arXiv preprint arXiv:1805.04807 , 2018

  4. [12]

    Adversarial robustness without adversarial training: A teacher-guided curriculum learning approach,

    A. Sarkar, A. Sarkar, S. Gali, and V . N Balasubramanian, “Adversarial robustness without adversarial training: A teacher-guided curriculum learning approach,” Advances in Neural Information Processing Systems, vol. 34, pp. 12 836–12 848, 2021

  5. [13]

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

    A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? the kitti vision benchmark suite,” in 2012 IEEE conference on computer vision and pattern recognition . IEEE, 2012, pp. 3354–3361

  6. [14]

    Argoverse: 3d tracking and forecasting with rich maps,

    M.-F. Chang, J. Lambert, P. Sangkloy, J. Singh, S. Bak, A. Hartnett, D. Wang, P. Carr, S. Lucey, D. Ramanan et al., “Argoverse: 3d tracking and forecasting with rich maps,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 8748– 8757

  7. [15]

    3d shapenets: A deep representation for volumetric shapes,

    Z. Wu, S. Song, A. Khosla, F. Yu, L. Zhang, X. Tang, and J. Xiao, “3d shapenets: A deep representation for volumetric shapes,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2015, pp. 1912–1920

  8. [16]

    Lightweight, uncertainty-aware conformalized visual odometry,

    A. C. Stutts, D. Erricolo, T. Tulabandhula, and A. R. Trivedi, “Lightweight, uncertainty-aware conformalized visual odometry,” in 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023, pp. 7742–7749

  9. [17]

    Starnet: Sensor trustworthiness and anomaly recognition via approx- imated likelihood regret for robust edge autonomy,

    N. Darabi, S. Tayebati, S. S., S. Ravi, T. Tulabandhula, and A. R. Trivedi, “Starnet: Sensor trustworthiness and anomaly recognition via approx- imated likelihood regret for robust edge autonomy,” arXiv preprint arXiv:2309.11006, 2023

  10. [18]

    Mutual information-calibrated conformal feature fusion for uncertainty- aware multimodal 3d object detection at the edge,

    A. C. Stutts, D. Erricolo, S. Ravi, T. Tulabandhula, and A. R. Trivedi, “Mutual information-calibrated conformal feature fusion for uncertainty- aware multimodal 3d object detection at the edge,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2...

  11. [19]

    Deep learning on 3d point clouds,

    S. A. Bello, S. Yu, C. Wang, J. M. Adam, and J. Li, “Deep learning on 3d point clouds,” Remote Sensing, vol. 12, no. 11, p. 1729, 2020

  12. [20]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space,

    C. R. Qi, L. Yi, H. Su, and L. J. Guibas, “Pointnet++: Deep hierarchical feature learning on point sets in a metric space,” Advances in neural information processing systems , vol. 30, 2017

  13. [21]

    Dynamic graph cnn for learning on point clouds,

    Y . Wang, Y . Sun, Z. Liu, S. E. Sarma, M. M. Bronstein, and J. M. Solomon, “Dynamic graph cnn for learning on point clouds,” ACM Transactions on Graphics (tog) , vol. 38, no. 5, pp. 1–12, 2019

  14. [22]

    Point transformer,

    H. Zhao, L. Jiang, J. Jia, P. H. Torr, and V . Koltun, “Point transformer,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 16 259–16 268

  15. [23]

    Conformalized multimodal uncertainty regression and reasoning,

    D. Parente, N. Darabi, A. C. Stutts, T. Tulabandhula, and A. R. Trivedi, “Conformalized multimodal uncertainty regression and reasoning,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 6985–6989

  16. [24]

    Enhancing 3d robotic vision robustness by minimizing adversarial mutual information through a curriculum training approach,

    N. Darabi, D. Jayasuriya, D. Naik, T. Tulabandhula, and A. R. Trivedi, “Enhancing 3d robotic vision robustness by minimizing adversarial mutual information through a curriculum training approach,” arXiv preprint arXiv:2409.12379, 2024

  17. [25]

    Sense less, generate more: Pre-training lidar perception with masked autoencoders for ultra- efficient 3d sensing,

    S. Tayebati, T. Tulabandhula, and A. R. Trivedi, “Sense less, generate more: Pre-training lidar perception with masked autoencoders for ultra- efficient 3d sensing,” arXiv preprint arXiv:2406.07833 , 2024

  18. [26]

    Navigating the unknown: Uncertainty-aware compute- in-memory autonomy of edge robotics,

    N. Darabi, P. Shukla, D. Jayasuriya, D. Kumar, A. C. Stutts, and A. R. Trivedi, “Navigating the unknown: Uncertainty-aware compute- in-memory autonomy of edge robotics,” in 2024 Design, Automation & Test in Europe Conference & Exhibition (DATE). IEEE, 2024, pp. 1–6

  19. [27]

    Robustness of 3d deep learning in an adversarial setting,

    M. Wicker and M. Kwiatkowska, “Robustness of 3d deep learning in an adversarial setting,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2019, pp. 11 767–11 775

  20. [28]

    Enhancing generalization of first-order meta-learning,

    M. Jayathilaka, “Enhancing generalization of first-order meta-learning,” 2019

  21. [29]

    On episodes, prototypical networks, and few-shot learning,

    S. Laenen and L. Bertinetto, “On episodes, prototypical networks, and few-shot learning,” Advances in Neural Information Processing Systems, vol. 34, pp. 24 581–24 592, 2021

  22. [30]

    Pointpillars: Fast encoders for object detection from point clouds,

    A. H. Lang, S. V ora, H. Caesar, L. Zhou, J. Yang, and O. Beijbom, “Pointpillars: Fast encoders for object detection from point clouds,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 12 697–12 705

  23. [31]

    Second: Sparsely embedded convolutional detection,

    Y . Yan, Y . Mao, and B. Li, “Second: Sparsely embedded convolutional detection,” Sensors, vol. 18, no. 10, p. 3337, 2018

  24. [32]

    Centerformer: Center-based transformer for 3d object detection,

    Z. Zhou, X. Zhao, Y . Wang, P. Wang, and H. Foroosh, “Centerformer: Center-based transformer for 3d object detection,” in European Confer- ence on Computer Vision . Springer, 2022, pp. 496–513

  25. [33]

    Transformation- equivariant 3d object detection for autonomous driving,

    H. Wu, C. Wen, W. Li, R. Yang, and C. Wang, “Transformation- equivariant 3d object detection for autonomous driving,” in AAAI, 2023

  26. [34]

    Pointnet: Deep learning on point sets for 3d classification and segmentation,

    C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmentation,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 652–660

  27. [35]

    Pointconv: Deep convolutional networks on 3d point clouds,

    W. Wu, Z. Qi, and L. Fuxin, “Pointconv: Deep convolutional networks on 3d point clouds,” in Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition , 2019, pp. 9621–9630

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.