Pith. sign in

REVIEW 3 major objections 5 minor 4 cited by

Factorized Implicit Global Convolution for Automotive Computational Fluid Dynamics Prediction

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read A factorized implicit global convolution network predicts automotive drag and surface pressure from large car meshes at quadratic complexity, outperforming previous 3D neural CFD methods on DrivAerNet and Ahmed body.

desk verdict Plausible architecture with strong reported results, but the core reparameterization is mis-derived and the headline numbers don't match the tables; worth a rigorous peer review, not acceptance as is. read the letter →

arxiv 2502.04317 v1 pith:UM4ORRCF submitted 2025-02-06 cs.CV

Chris Choy , Alexey Kamenev , Jean Kossaifi , Max Rietmann , Jan Kautz , Kamyar Azizzadenesheli This is my paper · ORCID
classification cs.CV
keywords computationalfluiddynamicsdragpredictionfactorizedimplicitgridsglobalconvolutionpointcloudpressureDrivAerNetAhmedbody
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 automotive CFD prediction, which normally requires processing huge 3D grids at cubic cost, can be done at quadratic cost by representing the grid as three factorized implicit grids, each with one low-resolution axis, and applying global convolutions in parallel. The result is a network that predicts per-face pressure and drag coefficient directly from car meshes. On DrivAerNet it reports a drag coefficient R2 of 0.957 and on Ahmed body a normalized pressure error of 0.89%, roughly a 70% reduction in absolute mean squared error over prior state-of-the-art methods. This matters because industrial meshes contain millions of vertices, and cubic-scale methods cannot process them at full resolution.

What carries the argument

The central mechanism is the Factorized Implicit Grid: a high-resolution 3D feature grid is decomposed into several grids, each with one axis at a low resolution r (typically r < 10) while the other two axes retain full resolution. A 3D convolution on the original grid is approximated by running global convolutions on each factorized grid in parallel and fusing the results. The 2D reparameterization flattens the low-resolution axis into the channel dimension, which lets a large kernel act globally along that axis while the computation is performed as a 2D convolution, giving quadratic scaling.

What would settle it

Take a random 3D feature tensor and a fixed kernel; compute the output of the reparameterized 2D convolution described in Section 3.3 and the output of a direct 3D convolution with the same kernel. If the outputs differ by more than floating-point round-off, the claimed equivalence fails. Independently, measure runtime and memory on grids with growing N for the same network; scaling beyond quadratic would falsify the complexity claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that a 3D convolution over a high-resolution domain can be approximated by three parallel convolutions on factorized grids, each with one low-resolution axis, and that when the kernel is large enough, the low-resolution axis can be flattened into the channel dimension to turn the 3D convolution into a 2D convolution with a global kernel. This reduces complexity from O($N^{3}$) to O($N^{2}$). The network combines these factorized implicit global convolutions in a U-shaped encoder-decoder, fuses the factorized grids after each convolution, and is trained jointly on drag and per-face pressure. The paper reports state-of-the-art results on DrivAerNet, with drag R2 of 0.957 and mean squared error 3.225E-5, and on the Ahmed body dataset, with a normalized pressure error of 0.89%.

Load-bearing premise

The load-bearing premise is that flattening the low-resolution axis into the channel dimension yields a convolution exactly equivalent to the original 3D convolution; if that algebraic equivalence is not exact, the quadratic-complexity claim and the global-convolution interpretation are not supported as stated.

Editorial extensions

If this is right

  • The quadratic complexity makes it feasible to run full-resolution neural CFD on meshes with millions of vertices on a single GPU, which cubic methods cannot do.
  • Jointly supervising drag and per-face pressure gives the encoder a dense training signal, improving drag accuracy beyond what drag-only regression achieves.
  • Larger convolution kernels, up to global size, improve pressure accuracy, and the 2D reparameterization keeps the added cost small enough to be worthwhile.
  • The factorized-grid approach is robust to the number of sampled input points, so it can work with variable mesh densities without retuning.

Reading between the lines

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

  • If the factorization is as faithful as the reported results suggest, the same O(N^2) recipe could be carried over to other dense 3D prediction problems, such as full flow-field regression or structural stress prediction, whenever the domain has one axis that can be kept low-resolution.
  • The algebraic equivalence in Section 3.3 is worth testing in isolation: replacing the direct 3D convolution with the flattened 2D convolution on identical weights should give equivalent outputs; an independent check would confirm the speedup is not buying accuracy through a different operation.
  • Because the network regresses drag directly without physics constraints, extrapolation to car geometries outside the training distribution may be less reliable than a solver-constrained model; adding physics-based losses is a natural extension the authors name as future work.
  • The Ahmed body comparison is reported as normalized per-face pressure error, not full flow fields; extending the evaluation to velocities or pressures off the surface would clarify how much of the global flow the factorized representation captures.
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 / 5 minor

Summary. The paper proposes FIGConvNet, a U-shaped network that predicts drag coefficients and surface pressure fields on automotive 3D meshes. The core idea is to represent a high-resolution 3D volume implicitly by a set of low-rank factorized grids, each with one low-resolution axis, and to apply large-kernel convolutions on these grids in parallel. The authors claim O(N^2) complexity instead of the O(N^3) of prior volumetric approaches, and report strong empirical results: R^2 = 0.957 for drag prediction on DrivAerNet and 0.89% normalized pressure error on the Ahmed body. The paper includes ablation studies on kernel size, grid rank, and fusion, and provides network configuration files in the appendix.

Significance. If the technical claims hold, the factorized implicit grid representation is a promising step toward making neural CFD practical for million-vertex automotive meshes. The reported speed and accuracy on DrivAerNet are compelling, and the paper gives enough architectural detail to reproduce the framework. However, the central derivation in Section 3.3 is not correct as written, and several headline numbers in the abstract and the Ahmed body section are inconsistent with the paper's own tables. These issues are load-bearing for the main claims and must be fixed before the manuscript can be accepted.

major comments (3)
  1. [3.3, Eq. (5)] The 2D reparameterization as displayed in Eq. (5) is not equivalent to the 3D convolution in Eq. (4). Writing the flattened index as s = k'·C + cin, the weight must depend on floor(s/C) to reproduce the kernel shift along the depth axis; the correct flattened weight is Wm(i',j',s,co) = W(i',j',floor(s/C), s mod C, co). Eq. (5) instead uses Wm(i',j',s mod C, co), which drops the depth-shift index and effectively sums C identical copies of a depth-averaged input. The sentence after Eq. (5) says both X and W are flattened, which would require Wm(i',j',s,co), contradicting the displayed formula. Since no code is released, the reader cannot tell whether the implementation uses the corrected kernel or the equation as written. This equivalence is load-bearing for the 'global convolution' and complexity claims, so please fix Eq. (5) and, if possible, provide a numerical equivalence check or release the code.
  2. [Abstract and Table 1] The abstract's claim of '40% improvement in relative mean squared error and a 70% improvement in absolute mean squared error over previous methods' is not supported by Table 1. The best baseline in Table 1 is PointNeXt with cd MSE = 4.577e-5; FIGConvNet achieves 3.225e-5, which is a 29.5% relative reduction. No row in Table 1 yields 40% or 70% under the standard definition of relative MSE improvement. Please correct the abstract and specify the baseline and metric used for each percentage.
  3. [5.3, Tables 5 and 6] The Ahmed body headline result of 0.89% normalized pressure error with a 68.29 MB model is not connected to any configuration in the controlled experiment of Table 6. Table 6 reports best pressure errors of 1.65% (at 140x42x45, kernel 9) and model sizes of at least 105 MB, with no configuration matching 0.89% or 68.29 MB. The manuscript does not state which grid resolution, rank, kernel size, or training schedule produces the 0.89% number, so the claimed order-of-magnitude improvement over GINO (9.01%) cannot be verified from the paper's own ablations. Please provide the exact configuration for Table 5 and reconcile the two tables.
minor comments (5)
  1. [3.3, Table 2 caption] The condition 'K ≥ 2r−' is incomplete; it should read 'K ≥ 2r−1' and should clarify whether r refers to the rank tuple or to each component of the rank.
  2. [3.1, Eq. (3)] The product notation ∏_m Y_m is not well-defined for factorized grids of different resolutions; Section 3.4 describes fusion via trilinear interpolation and aggregation, so Eq. (3) should be aligned with the actual aggregation operation.
  3. [5.2, Table 1] The statement that all baselines except DrivAerNet DGCNN were trained with both pressure and drag losses is not visible in Table 1; add a footnote to the table so that the comparison rules are transparent.
  4. [Appendix B] The sentence 'The dataset is proprietary from NVIDIA Corp.' conflicts with the description of DrivAerNet as a public dataset; clarify which parts of the data are proprietary and what the release status is.
  5. [5.3] The pressure error metric is described as 'normalized L2 pressure error per vertex' in Table 5 but simply 'Pressure Error' in Table 6; define the metric once and use it consistently in both tables.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: held-out test predictions are genuine and the O(N^2) derivation is not self-referential; the score reflects minor baseline self-citation, test-set hyperparameter selection, and an unverified Eq. 5 equivalence.

full rationale

The paper's predictive claims are evaluated on held-out test splits (the official DrivAerNet test set and the 10% Ahmed-body test split), so the drag R2=0.957 and 0.89% pressure error are predictions rather than fits to the target. The factorized implicit grid construction (Eqs. 1-3) is a representation approximation, and the O(N^2) complexity follows from bounding the rank r and the number M of factorized grids, not from the benchmark targets. Self-citations such as GINO (Li et al. 2023, with overlapping authors) and Kossaifi et al. 2023/2024 are used as baselines or related work and are not load-bearing for the architecture derivation. Two caveats do not rise to circularity under the definitions in this review. First, Section 3.3, Eq. 5 writes the flattened kernel as Wm(i',j',s mod C,co), dropping the floor(s/C) depth-tap index, so the displayed equivalence to 3D convolution is not established as written; since no code is released (appendix B.3 says code will be released upon acceptance), the O(N^2) global-convolution claim depends on an unverified implementation detail. This is a correctness/reproducibility gap, not a self-referential reduction. Second, Tables 2 and 3 select kernel size and rank using the official DrivAerNet test set and then report final test-set numbers in Table 1, which biases the benchmark comparison but does not make the predictions equal to the fitted hyperparameters by construction. Score 2 reflects these minor caveats rather than load-bearing circularity.

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

The central claim rests on the expressive power of factorized implicit grids, the exactness of the 2D reparameterization, and the representativeness of two simulated datasets. The rank and kernel size are explicit free parameters chosen by experiment. The FIG representation is a new model component with empirical support but no formal approximation guarantees.

free parameters (4)
  • rank r=(rx,ry,rz) of factorized grids = Explored as (4,4,4), (3,2,2), (5,3,2), (10,6,4), (10,10,10); config uses (5,3,2)-like resolutions 5x150x100…
    Controls the size of the low-resolution axis in each factorized grid and directly affects accuracy and inference time (Tables 3 and 4).
  • convolution kernel size K = Main config uses 5; swept over 3, 5, 7, 9, 11 in Tables 2 and 6
    Kernel size must reach K >= 2r-1 to be 'global'; larger kernels improve accuracy but slow inference, so the choice is empirical.
  • network width and depth = hidden_channels [16, 32, 48], num_levels 2, num_down_blocks [1,1], num_up_blocks [1,1]
    Architecture capacity hyperparameters chosen by hand without ablation, affecting model size and accuracy.
  • ellipsoid covariance Sigma for point convolution = Not specified in the paper
    Defines the neighborhood in Eq. 8 for converting point clouds to factorized grids; the shape and scale are not reported, making the input encoding under-specified.
assumptions (4)
  • domain assumption Factorized implicit grids with product decoding (Eq. 1-2) can approximate the high-resolution explicit grid X well enough for pressure and drag prediction.
    Assumed in Sections 3.1-3.2; no approximation error bound is given.
  • ad hoc to paper The 2D reparameterization of the 3D convolution is exactly equivalent to the original 3D convolution (Eq. 4-7).
    Asserted in Section 3.3, but the derivation appears to drop the kernel shift index, so the equivalence is not established.
  • domain assumption The U-Net with joint pressure and drag losses provides sufficient supervision for drag prediction given the small dataset size.
    Section 3.6 motivates the auxiliary pressure task, but there is no ablation showing it is necessary or sufficient.
  • domain assumption The DrivAerNet and Ahmed body datasets are representative of automotive CFD for the generalization claims.
    The paper generalizes to 'automotive CFD' from two simulated datasets, neither of which is validated against extensive wind-tunnel data.
invented entities (1)
  • Factorized implicit grid (FIG) representation independent evidence
    purpose: Compact encoding of a high-resolution 3D field as M grids, each with one low-resolution axis, decoded by MLPs (Eq. 1-2).
    Its adequacy is tested by held-out pressure and drag prediction on DrivAerNet and Ahmed body test sets, so it has an empirical handle, though no theoretical guarantee is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Factorized Implicit Global Convolution for Automotive Computational Fluid Dynamics Prediction." pith.science (2026). https://pith.science/paper/UM4ORRCF

@misc{pith2026250204317,
  author       = {Pith},
  title        = {Pith review of: Factorized Implicit Global Convolution for Automotive Computational Fluid Dynamics Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UM4ORRCF}},
  note         = {Machine review of arXiv:2502.04317}
}
abstract

Computational Fluid Dynamics (CFD) is crucial for automotive design, requiring the analysis of large 3D point clouds to study how vehicle geometry affects pressure fields and drag forces. However, existing deep learning approaches for CFD struggle with the computational complexity of processing high-resolution 3D data. We propose Factorized Implicit Global Convolution (FIGConv), a novel architecture that efficiently solves CFD problems for very large 3D meshes with arbitrary input and output geometries. FIGConv achieves quadratic complexity $O(N^2)$, a significant improvement over existing 3D neural CFD models that require cubic complexity $O(N^3)$. Our approach combines Factorized Implicit Grids to approximate high-resolution domains, efficient global convolutions through 2D reparameterization, and a U-shaped architecture for effective information gathering and integration. We validate our approach on the industry-standard Ahmed body dataset and the large-scale DrivAerNet dataset. In DrivAerNet, our model achieves an $R^2$ value of 0.95 for drag prediction, outperforming the previous state-of-the-art by a significant margin. This represents a 40% improvement in relative mean squared error and a 70% improvement in absolute mean squared error over previous methods.

Figures

Figures reproduced from arXiv: 2502.04317 by the authors.

Figure 1
Figure 1. FIGConvNet: ConvNet for drag prediction using FIG convolution blocks. The encoder and decoder consist of a set of FIG convolution blocks and we connect the encoder and decoder with skip connections. The output of the encoder is used for drag prediction and the output of the decoder is used for pressure prediction. 3 Factorized Implicit Global ConvNet In this section, we introduce our factorized implicit global convo… view at source ↗
Figure 2
Figure 2. From left to right, we have a regular convolution, a separable convolution, and our proposed [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Factorized Implicit Global Convolution 3D: The FIG convolution first creates a set of voxel grids that factorizes the domain. This allows representing a high resolution voxel grid domain implicitly that can be computationally prohibitive to save explicitly. Then, a set of global convolution operations are applied in parallel to these voxel grids to capture the global context. Finally, the voxel grids are aggregated … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Point Convolution: The features from source and target nodes as well as offset are fed into an MLP to lift the features, which are then aggregated and projected back to the original feature space using an MLP. We discussed how we perform global convolution on the facto…
Figure 6
Figure 6. Figure 6: Normalized Pressure Prediction and Error Visualization on DrivAerNet. Our network predicts both drag coefficients and per vertex pressure. We visualize the ground truth pressure and prediction along with the absolute error of the pressure. Note that the pressures are n…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. A Benchmarking Framework for AI models in Automotive Aerodynamics

    cs.LG 2025-07 conditional novelty 6.0 of 10

    A new benchmarking framework standardizes evaluation of AI automotive aerodynamics models, demonstrated on three models with the DrivAerML dataset.

  2. A Mixture of Experts Gating Network for Enhanced Surrogate Modeling in External Aerodynamics

    cs.LG 2025-08 conditional novelty 5.0 of 10

    A mixture-of-experts gating network that fuses predictions from DoMINO, X-MeshGraphNet, and FigConvNet reduces L-2 prediction error for automotive surface pressure and wall shear stress below each individual expert on...

  3. GeoTransolver: Learning Physics on Irregular Domains Using Multi-scale Geometry Aware Physics Attention Transformer

    cs.LG 2025-12 conditional novelty 4.0 of 10

    GeoTransolver, a geometry-aware attention transformer, improves surrogate CFD accuracy over existing baselines on three automotive/aerospace datasets, but the paper has major reporting gaps.

  4. Learning Mappings in Mesh-based Simulations

    cs.LG 2025-06 conditional novelty 3.0 of 10

    A bilinear scatter encoding plus a masked UNet yields competitive surrogate accuracy and data efficiency on several mesh-based simulation benchmarks, though the encoding is a standard technique.

Reference graph

Works this paper leans on

19 extracted references · 10 canonical work pages · cited by 4 Pith papers

  1. [2]

    3.1) e.g

    (Sec. 3.1) e.g. The three grid resolutions we used for the first three rows are 6×280×180, 560×2×180, 560×208×2. Max Resolution Kernel Size Pressure Error Model Size (MB) 560×208×180 3 3.40% 105.0 7 3.31% 417.1 11 2.56% 979.7 280×104×90 3 2.89% 105.0 7 3.05% 417.1 11 2.93% 979.7 140×42×45 9 1.65% 667.29 11 2.59% 979.7 DrivAerNet datasets is the parametric...

    work page 2012

  2. [9]

    Deep learning for real-time aerodynamic evaluations of arbitrary vehicle shapes

    Sam Jacob Jacob, Markus Mrosek, Carsten Othmer, and Harald Köstler. Deep learning for real-time aerodynamic evaluations of arbitrary vehicle shapes. arXiv preprint arXiv:2108.05798,

    arXiv

  3. [13]

    Fourier neural operator for parametric partial differential equations

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. arXiv preprint arXiv:2010.08895, 2020a. Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkum...

    arXiv 2010

  4. [14]

    Kovachki, Chris Choy, Boyi Li, Jean Kossaifi, Shourya Prakash Otta, Mo- hammad Amin Nabian, Maximilian Stadler, Christian Hundt, Kamyar Azizzadenesheli, et al

    Zongyi Li, Nikola B. Kovachki, Chris Choy, Boyi Li, Jean Kossaifi, Shourya Prakash Otta, Mo- hammad Amin Nabian, Maximilian Stadler, Christian Hundt, Kamyar Azizzadenesheli, et al. Geometry-informed neural operator for large-scale 3d pdes. arXiv preprint arXiv:2309.00583,

    arXiv

  5. [15]

    Jaideep Pathak, Shashank Subramanian, Peter Harrington, Sanjeev Raja, Ashesh Chattopadhyay, Morteza Mardani, Thorsten Kurth, David Hall, Zongyi Li, Kamyar Azizzadenesheli, et al

    doi: 10.1109/JPROC.2021.3074329. Jaideep Pathak, Shashank Subramanian, Peter Harrington, Sanjeev Raja, Ashesh Chattopadhyay, Morteza Mardani, Thorsten Kurth, David Hall, Zongyi Li, Kamyar Azizzadenesheli, et al. Fourcast- net: A global data-driven high-resolution weather model using adaptive fourier neural operators. arXiv preprint arXiv:2202.11214,

    work page arXiv 2021

  6. [16]

    Learning mesh- based simulation with graph networks

    Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter W Battaglia. Learning mesh- based simulation with graph networks. arXiv preprint arXiv:2010.03409, 2020a. Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter W. Battaglia. Learning mesh- based simulation with graph networks. arXiv preprint arXiv:2010.03409, 2020b. Charles R Qi...

    arXiv 2010

  7. [18]

    14 A Appendix B Datasets The foundation of CFD in the automotive industry provides insight into design and engineering. The comprehensive previous texts provide a solid overview of computational methods in fluid dynamics and dedicate a comprehensive overview of traditional CFD techniques [Ferziger et al., 2019] along with specification in automotive aerod...

    work page 2019

  8. [20]

    for the baseline implementation and, with configuration, you can specify the network architecture. B.2 Baseline Implementations We use the OpenPoint, an open-soruce 3D point cloud library [Qian et al., 2022] to implement Point- Net++ [Qi et al., 2017b], DeepGCN [Li et al., 2019], AssaNet [Qian et al., 2021], PointNeXt [Qian et al., 2022], and PointBERT [Y...

    work page 2022

Show all 19 references
  1. [1984]

    Cp-decomposition with tensor power method for convolutional neural networks compression

    Marcella Astrid and Seung-Ik Lee. Cp-decomposition with tensor power method for convolutional neural networks compression. CoRR, abs/1701.07148,

    arXiv

  2. [2012]

    Hermosilla, T

    P. Hermosilla, T. Ritschel, P-P Vazquez, A. Vinacua, and T. Ropinski. Monte carlo convolution for learning on non-uniformly sampled point clouds. ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia 2018),

    2018

  3. [2013]

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox

    doi: 10.1109/CVPR.2013.355. Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5...

    2013 doi

  4. [2016]

    Kossaifi, A

    J. Kossaifi, A. Toisoul, A. Bulat, Y . Panagakis, T. M. Hospedales, and M. Pantic. Factorized higher-order cnns with an application to spatio-temporal emotion estimation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6059–6068, Los Alamit...

    2020

  5. [2017]

    Point convolutional neural networks by extension operators

    Matan Atzmon, Haggai Maron, and Yaron Lipman. Point convolutional neural networks by extension operators. arXiv preprint arXiv:1803.10091,

    arXiv

  6. [2019]

    Drivaernet: A parametric car dataset for data- driven aerodynamic design and graph-based drag prediction

    Mohamed Elrefaie, Angela Dai, and Faez Ahmed. Drivaernet: A parametric car dataset for data- driven aerodynamic design and graph-based drag prediction. arXiv preprint arXiv:2403.08055,

    arXiv

  7. [2020]

    doi: 10.1109/CVPR42600.2020.00610

    IEEE Computer Society. doi: 10.1109/CVPR42600.2020.00610. URL https://doi.ieeecomputersociety.org/10.1109/CVPR42600.2020.00610. Jean Kossaifi, Nikola Kovachki, Kamyar Azizzadenesheli, and Anima Anandkumar. Multi-grid tensorized fourier neural operator for high-resolution pdes,

    2020

  8. [2021]

    Tianjin Huang, Lu Yin, Zhenyu Zhang, Li Shen, Meng Fang, Mykola Pechenizkiy, Zhangyang Wang, and Shiwei Liu

    URL https://arxiv.org/abs/2106.09685. Tianjin Huang, Lu Yin, Zhenyu Zhang, Li Shen, Meng Fang, Mykola Pechenizkiy, Zhangyang Wang, and Shiwei Liu. Are large kernels better teachers than transformers for convnets? In International Conference on Machine Learning, pages 14023–140...

    arXiv

  9. [2022]

    Shapenet: An information-rich 3d model repository

    Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012,

    arXiv

  10. [2023]

    Submanifold sparse convolutional networks

    Benjamin Graham and Laurens van der Maaten. Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307,

    arXiv

  11. [2024]

    URL https://openreview.net/forum?id= AWiDlO63bH

    ISSN 2835-8856. URL https://openreview.net/forum?id= AWiDlO63bH. Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Alexan- der Pritzel, Suman Ravuri, Timo Ewalds, Ferran Alet, Zach Eaton-Rosen, et al. Graphcast: Learning skillful medium-ra...

    arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.