Back

Multi-view Convolutional Neural Networks for 3D Shape Recognition

Hang SuSubhransu MajiEvangelos KalogerakisErik Learned-Miller

article2015ICCV3,660 citations

Introduces a multi-view convolutional neural network architecture that pools standard 2D rendered views into a compact descriptor, demonstrating that 2D image representations outperform native 3D voxel and mesh models on 3D shape and sketch recognition.

Listen

A longstanding challenge in computer vision is how to represent three-dimensional shapes for accurate recognition when only two-dimensional images are typically available. Large online repositories now supply thousands of 3D models, yet prior classifiers built directly on voxel or mesh representations have achieved only modest accuracy. The paper examines whether collections of 2D renderings can serve as a stronger basis for recognition than native 3D descriptors.

The work set out to measure the recognition performance obtainable from rendered views alone and to develop a compact descriptor that fuses information across multiple views. Researchers rendered twelve fixed viewpoints of each model in the ModelNet40 benchmark, passed the images through convolutional networks pre-trained on ImageNet and fine-tuned on the rendered views, and compared results against established 3D shape descriptors and against simpler view-averaging baselines.

A single rendered view classified by a fine-tuned network already reached 84 percent accuracy, exceeding the previous best 3D method by roughly seven points. Supplying all twelve views raised accuracy to 87–89 percent and mean average precision for retrieval to 69 percent; a learned low-rank metric further lifted retrieval to 79 percent while shrinking the descriptor to 128 dimensions. The same multi-view architecture improved sketch recognition from 79 percent to 87 percent, approaching human performance on cleaned data, and enabled direct sketch-to-model retrieval at 36 percent mAP without additional domain-specific training.

These gains indicate that high-resolution two-dimensional projections, combined with existing image networks, capture category-discriminating structure more effectively than coarse voxel grids. The resulting descriptors are compact, fast to compare, and immediately usable for both classification and retrieval tasks.

Further gains will likely require systematic study of viewpoint selection and the extension of the same aggregation approach to video or real-world multi-view captures. The current results rest on upright-oriented synthetic models rendered under controlled lighting; performance on unoriented or noisy scans remains untested. Overall the evidence strongly supports view-based convolutional pipelines for 3D shape tasks, with the largest practical advantages appearing in retrieval and cross-domain search.

University of Massachusetts Amherstcomputer visionconvolutional neural networksmultiple viewsMulti-view Convolutional Neural Networks3D shape recognitionshape retrieval
arXiv: 1505.00880suhangpro/mvcnn
  • Paper: An Introduction to Convolutional Neural Networks, Keiron O'Shea et al. (2015). Reading this foundational tutorial on convolutional neural networks provides the core architectural background necessary to understand how the source applies CNNs to 2D views of 3D objects.
  • Paper: CNN Features Off-the-Shelf: An Astounding Baseline for Recognition, Ali Sharif Razavian et al. (2014). This paper establishes the powerful baseline of using pre-trained image classification CNNs off-the-shelf, which directly motivates the source's investigation into applying standard CNN architectures to rendered 2D views.
Cover for Multi-view Convolutional Neural Networks for 3D Shape Recognition

Abstract

A longstanding question in computer vision concerns the representation of 3D shapes for recognition: should 3D shapes be represented with descriptors operating on their native 3D formats, such as voxel grid or polygon mesh, or can they be effectively represented with view-based descriptors? We address this question in the context of learning to recognize 3D shapes from a collection of their rendered views on 2D images. We first present a standard CNN architecture trained to recognize the shapes' rendered views independently of each other, and show that a 3D shape can be recognized even from a single view at an accuracy far higher than using state-of-the-art 3D shape descriptors. Recognition rates further increase when multiple views of the shapes are provided. In addition, we present a novel CNN architecture that combines information from multiple views of a 3D shape into a single and compact shape descriptor offering even better recognition performance. The same architecture can be applied to accurately recognize human hand-drawn sketches of shapes. We conclude that a collection of 2D views can be highly informative for 3D shape recognition and is amenable to emerging CNN architectures and their derivatives.

Topics

computer visionconvolutional neural networksmultiple viewsshape descriptorsviewpointarchitecture designsfeature descriptorstwo dimensional

Keywords

Multi-view Convolutional Neural Networks3D shape recognitionshape retrievalview-based descriptorsvoxel-based representationshape representationsketch recognition

Table of Contents

  • 1. Introduction
  • 2. Related Work
  • 3. Method
  • 3.1. Input: A multi-view representation
  • 3.2. Recognition with the multi-view representation
  • 3.3. Multi-view CNN: Learning to aggregate views
  • 4. Experiments
  • 4.1. 3D shape classification and retrieval
  • 4.2. Sketch recognition: jittering revisited
  • 4.3. Sketch-based 3D shape retrieval
  • 5. Conclusion
  • A. 3D shape classification
  • B. Image-specific class saliency visualization across views
  • C. Sketch classification
  • D. Document changelog
  • References

Knowls

  1. Knowl 1 — Multi-View Convolutional Neural Network Architecture

    model/method

    The Multi-View Convolutional Neural Network (MVCNN) is designed to aggregate information from a set of rendered 2D views of a 3D shape into a single, compact descriptor.

    The network comprises three stages:

    1. First Branch Pipeline (CNN1\mathrm{CNN}_1): Each 2D rendered image of a 3D shape is passed independently through the initial layers of a convolutional network. All branches share the exact same weights.
    2. View-Pooling Layer: Feature representations across all KK view branches are aggregated using an element-wise maximum operation across the view dimension. For activations zk\mathbf{z}_k from view k∈{1,…,K}k \in \{1, \dots, K\}, the pooled activation zpool\mathbf{z}_{\text{pool}} has entries:
    zpool,j=max⁡k=1,…,Kzk,jz_{\text{pool}, j} = \max_{k=1,\dots,K} z_{k, j}

    Element-wise max-pooling across views allows the network to select the most discriminative features across viewpoints while discarding uninformative views, outperforming element-wise mean-pooling. 3. Second Pipeline (CNN2\mathrm{CNN}_2): The aggregated feature map is passed through the remaining layers of the network (e.g., remaining fully connected layers and a softmax output layer).

    When using the VGG-M architecture, the base network consists of five convolutional layers (conv1\text{conv}_1 to conv5\text{conv}_5) followed by three fully connected layers (fc6\text{fc}_6, fc7\text{fc}_7, fc8\text{fc}_8). View-pooling is placed immediately after conv5\text{conv}_5. The 4096-dimensional output of the fc7\text{fc}_7 layer (after ReLU activation) serves as the compact 3D shape descriptor for downstream classification and retrieval tasks. The entire architecture forms a directed acyclic graph trainable end-to-end using stochastic gradient descent with back-propagation.

  2. Knowl 2 — 3D Shape Classification and Retrieval Performance on ModelNet40

    data/table

    The Princeton ModelNet40 dataset contains 12,311 3D CAD models from 40 categories. The classification accuracy and retrieval mean Average Precision (mAP) are evaluated comparing native 3D descriptors (Spherical Harmonics, LightField, and 3D ShapeNets) against 2D view-based methods (Fisher Vectors on SIFT, standard CNNs using VGG-M, and MVCNN).

    Method Training Config. Test Config. Classification Retrieval
    Pre-train Fine-tune #Views #Views (Accuracy) (mAP)
    (1) SPH - - - - 68.2% 33.3%
    (2) LFD - - - - 75.5% 40.9%
    (3) 3D ShapeNets ModelNet40 ModelNet40 - - 77.3% 49.2%
    (4) FV - ModelNet40 12 1 79.3% 37.6%
    (5) FV, 12×\times - ModelNet40 12 12 85.1% 43.8%
    (6) CNN ImageNet1K - - 1 82.4% 43.3%
    (7) CNN, f.t. ImageNet1K ModelNet40 12 1 84.1% 61.0%
    (8) CNN, 12×\times ImageNet1K - - 12 86.4% 49.0%
    (9) CNN, f.t., 12×\times ImageNet1K ModelNet40 12 12 87.3% 62.1%
    (10) MVCNN, 12×\times ImageNet1K - - 12 86.4% 49.2%
    (11) MVCNN, f.t., 12×\times ImageNet1K ModelNet40 12 12 88.8% 69.4%
    (12) MVCNN, f.t.+metric, 12×\times ImageNet1K ModelNet40 12 12 88.9% 78.9%

    Key findings from these benchmarks include:

    • Even with a single rendered view at test time, 2D view-based approaches (Fisher Vector at 79.3% and fine-tuned CNN at 84.1%) outperform voxel-based 3D ShapeNets (77.3%).
    • Fine-tuned MVCNN using 12 views achieves 88.8% classification accuracy (an 11.5% absolute gain over 3D ShapeNets) and 69.4% retrieval mAP.
    • Adding low-rank Mahalanobis metric learning to MVCNN features dramatically increases retrieval mAP from 69.4% to 78.9%.

  3. Knowl 3 — Multi-View Rendering Pipeline for 3D Polygon Meshes

    model/method

    To convert a 3D polygon mesh into a 2D multi-view representation, 12 virtual camera viewpoints are arranged under the assumption that the 3D models are consistently upright oriented:

    1. Camera Placement: 12 virtual cameras are placed every 30∘30^\circ azimuthally around the mesh on a circle elevated by 30∘30^\circ above the ground plane.
    2. Camera Target: Each camera points towards the centroid of the mesh, defined as the area-weighted average of the face centers:
    c=∑fAfcf∑fAf\mathbf{c} = \frac{\sum_{f} A_f \mathbf{c}_f}{\sum_f A_f}

    where AfA_f is the surface area of triangular/polygonal face ff and cf\mathbf{c}_f is its geometric center. 3. Rendering & Illumination: Meshes are rendered under perspective projection and uniformly scaled to fit within the viewing volume. Polygons are illuminated using the Phong reflection model, interpolating reflected intensities across polygon vertices.

    Generating all 12 views requires under 10 ms per mesh on modern graphics hardware.

  4. Knowl 4 — Two-Way Average Minimum Distance for View-Based 3D Shape Retrieval

    equation

    For 3D shape retrieval using separate sets of 2D view descriptors without network-level pooling, the asymmetric distance from a single view image xi\mathbf{x}_i of shape x\mathbf{x} to the entire multi-view collection {yj}j=1ny\{\mathbf{y}_j\}_{j=1}^{n_y} of shape y\mathbf{y} is defined as:

    d(xi,y)=min⁡j∥xi−yj∥2d(\mathbf{x}_i, \mathbf{y}) = \min_{j} \|\mathbf{x}_i - \mathbf{y}_j\|_2

    To construct a symmetric distance measure between shape x\mathbf{x} (containing nxn_x view descriptors {xi}i=1nx⊂Rd\{\mathbf{x}_i\}_{i=1}^{n_x} \subset \mathbb{R}^d) and shape y\mathbf{y} (containing nyn_y view descriptors {yj}j=1ny⊂Rd\{\mathbf{y}_j\}_{j=1}^{n_y} \subset \mathbb{R}^d), distances are averaged in both directions:

    Dist(x,y)=12∑j=1nymin⁡i∥xi−yj∥2ny+12∑i=1nxmin⁡j∥xi−yj∥2nx\mathrm{Dist}(\mathbf{x}, \mathbf{y}) = \frac{1}{2} \frac{\sum_{j=1}^{n_y} \min_{i} \|\mathbf{x}_i - \mathbf{y}_j\|_2}{n_y} + \frac{1}{2} \frac{\sum_{i=1}^{n_x} \min_{j} \|\mathbf{x}_i - \mathbf{y}_j\|_2}{n_x}

    where ∥⋅∥2\|\cdot\|_2 denotes the Euclidean (L2L_2) norm. This two-way average minimum formulation outperforms both the global minimum pair distance min⁡i,j∥xi−yj∥2\min_{i,j} \|\mathbf{x}_i - \mathbf{y}_j\|_2 and the distance between mean view descriptors.

  5. Knowl 5 — Low-Rank Mahalanobis Metric Learning for MVCNN Descriptors

    model/method

    Because MVCNN fine-tuning optimizes a classification cross-entropy loss rather than an explicit ranking loss, retrieval performance is improved post-hoc by learning a low-rank Mahalanobis metric.

    A linear transformation matrix W∈Rp×dW \in \mathbb{R}^{p \times d} is learned via Large-Margin Metric Learning to project the dd-dimensional fc7\text{fc}_7 shape descriptor ϕ∈Rd\boldsymbol{\phi} \in \mathbb{R}^d into a lower-dimensional space Wϕ∈RpW\boldsymbol{\phi} \in \mathbb{R}^p (with d=4096d = 4096 and p=128p = 128). The projection objective minimizes L2L_2 distances between projected descriptors of the same category while maintaining a large margin between descriptors of different categories:

    DW(ϕa,ϕb)=∥Wϕa−Wϕb∥2\mathcal{D}_W(\boldsymbol{\phi}_a, \boldsymbol{\phi}_b) = \|W \boldsymbol{\phi}_a - W \boldsymbol{\phi}_b\|_2

    Applying this metric reduces descriptor storage dimensionality from 4096 to 128 and increases ModelNet40 retrieval mAP from 69.4% to 78.9% without diminishing classification accuracy (88.8% to 88.9%).

  6. Knowl 6 — Impact of View-Pooling Layer Location in MVCNN

    data/table

    The view-pooling layer in MVCNN can theoretically be inserted at any point in the underlying convolutional architecture. Classification accuracy and retrieval mAP on ModelNet40 across different layer placements within the VGG-M network are evaluated as follows:

    Layer Classification (Accuracy) Retrieval (mAP)
    conv3_3 85.9% 62.6%
    conv4_4 88.9% 65.8%
    conv5_5 88.8% 69.4%
    fc6_6 87.8% 68.3%
    fc7_7 87.5% 69.7%

    While overall performance is relatively robust across layers, placing view-pooling at conv5\text{conv}_5 achieves the best overall balance between classification accuracy (88.8%) and retrieval mAP (69.4%).

  7. Knowl 7 — Sketch Classification via Multi-View Feature Aggregation Over Jittered Inputs

    data/table

    MVCNN's view-aggregation architecture can be applied to 2D image domains by treating data jittering (rotations and reflections) as different "views" of a single sample. On the 160-category cleaned human sketch dataset (SketchClean), sketches are augmented with 6 jittered versions (in-plane rotations of −45∘,0∘,45∘-45^\circ, 0^\circ, 45^\circ combined with horizontal reflections):

    Method Augmentation Accuracy
    (1) FV - 79.0%
    (2) CNN M - 77.3%
    (3) CNN M, fine-tuned - 84.0%
    (4) CNN M, fine-tuned 6×\times 85.5%
    (5) MVCNN M, fine-tuned 6×\times 86.3%
    (6) CNN VD - 69.3%
    (7) CNN VD, fine-tuned - 86.3%
    (8) CNN VD, fine-tuned 6×\times 86.0%
    (9) MVCNN VD, fine-tuned 6×\times 87.2%
    (10) Human performance n/a 93.0%

    Using MVCNN to pool over the 6 jittered representations outperforms both standard test-time average pooling across jittered crops and standard single-image CNNs across both VGG-M and the 16-layer VGG-VD architecture, advancing state-of-the-art sketch recognition from 79.0% to 87.2% (approaching the 93.0% human benchmark).

  8. Knowl 8 — Cross-Domain Sketch-Based 3D Shape Retrieval Framework

    model/method

    To perform 3D shape retrieval directly from hand-drawn 2D sketches across domain disparities:

    1. 3D Mesh Domain Transformation: 3D CAD meshes are rendered into sketch-like line drawings by rendering depth buffers (zz-buffers) from 12 viewpoints elevated by 30∘30^\circ and running Canny edge detection on each depth image.
    2. Feature Extraction: Both the 12 depth edge maps per 3D model and 6 jittered variations (−45∘,0∘,45∘-45^\circ, 0^\circ, 45^\circ rotations and horizontal reflections) per query sketch are passed through an off-the-shelf ImageNet-trained VGG-M network to extract fc7\text{fc}_7 activation vectors.
    3. Ranking Metric: Shapes are ranked against the query sketch using the symmetric two-way average minimum Euclidean distance between the sets of feature vectors.

    On an evaluation benchmark of 193 query sketches and 790 CAD models spanning 10 overlapping classes between SketchClean and ModelNet40, this zero-shot cross-domain matching method achieves 36.1% mAP without any cross-domain fine-tuning.

  9. Knowl 9 — Multi-View Class Saliency Map Formulation

    equation

    For a 3D shape SS represented by a set of KK 2D rendered views {I1,I2,…,IK}\{I_1, I_2, \dots, I_K\}, the saliency of each pixel across views with respect to the network's class score FcF_c for ground truth class cc (e.g., from the fc8\text{fc}_8 layer) is defined as the gradient vector of FcF_c with respect to the input images:

    [w1,w2,…,wK]=[∂Fc∂I1∣S,∂Fc∂I2∣S,…,∂Fc∂IK∣S][w_1, w_2, \dots, w_K] = \left[ \left.\frac{\partial F_c}{\partial I_1}\right|_S, \left.\frac{\partial F_c}{\partial I_2}\right|_S, \dots, \left.\frac{\partial F_c}{\partial I_K}\right|_S \right]

    In MVCNN, [w1,…,wK][w_1, \dots, w_K] is computed via a single back-propagation pass with all network parameters fixed. Because gradients only propagate through the winning inputs of the view-pooling max operation, the resulting gradient energy ∥wk∥\|w_k\| directly identifies which views are canonical and informative for recognizing the category, as well as which local visual features within those views (such as bathtub faucets or desk handles) drive the classification decision.

Coverage note — Confusion matrix values and specific visual examples of misclassifications from the appendix were omitted as they detail qualitative analysis rather than core algorithmic or quantitative contributions.

References

  1. 1.The Princeton ModelNet. http://modelnet.cs.princeton.edu/. [Online; accessed March 2015]. 5
  2. 2.A. M. Bronstein, M. M. Bronstein, M. Ovsjanikov, and L. J. Guibas. Shape google: Geometric words and expressions for invariant shape retrieval. ACM Trans. Graphics, 30, 2011. 2
  3. 3.A. M. Bronstein, R. Litman, M. M. Bronstein, and I. Kokkinos. Intrinsic shape context descriptors for deformable shapes. In Proc. CVPR, 2012. 2
  4. 4.K. Chatfield, K. Simonyan, A. Vedaldi, and A. Zisserman. Return of the devil in the details: Delving deep into convolutional nets. In Proc. BMVC, 2014. 4, 6, 7
  5. 5.S. Chaudhuri and V. Koltun. Data-driven suggestions for creativity support in 3D modeling. ACM Trans. Graph, 29(6), 2010. 2
  6. 6.D.-Y. Chen, X.-P. Tian, Y.-T. Shen, and M. Ouhyoung. On visual similarity based 3D model retrieval. Proc. Eurographics, 22(3):223–232, 2003. 3, 4, 5
  7. 7.S. Chopra, R. Hadsell, and Y. LeCun. Learning a similarity metric discriminatively, with application to face verification. In Proc. CVPR, 2005. 8
  8. 8.M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proc. CVPR, 2014. 3
  9. 9.C. M. Cyr and B. B. Kimia. A similarity-based aspect-graph approach to 3D object recognition. 57(1), 2004. 3
  10. 10.J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proc. CVPR, 2009. 1, 2
  11. 11.J. Donahue, Y. Jia, O. Vinyals, J. Hoffman, N. Zhang, E. Tzeng, and T. Darrell. DeCAF: A deep convolutional activation feature for generic visual recognition. CoRR, abs/1310.1531, 2013. 3, 4
  12. 12.M. Eitz, J. Hays, and M. Alexa. How do humans sketch objects? ACM Trans. Graph. (Proc. SIGGRAPH), 31(4):44:1–44:10, 2012. 2, 7
  13. 13.M. Eitz, R. Richter, T. Boubekeur, K. Hildebrand, and M. Alexa. Sketch-based shape retrieval. ACM Trans. Graph., 31(4), 2012. 3, 7, 8
  14. 14.R. B. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proc. CVPR, 2014. 3
  15. 15.I. J. Goodfellow, D. Warde-Farley, M. Mirza, A. Courville, and Y. Bengio. Maxout networks. ArXiv e-prints, Feb. 2013. 4
  16. 16.B. K. P. Horn. Extended gaussian images. Proc. of the IEEE, 72(12):1671–1686, 1984. 2
  17. 17.M. Kazhdan, T. Funkhouser, and S. Rusinkiewicz. Rotation invariant spherical harmonic representation of 3D shape descriptors. Proc. Symposium of Geometry Processing, 2003. 2, 4, 5
  18. 18.J. Knopp, M. Prasad, G. Willems, R. Timofte, and L. Van Gool. Hough transform and 3D SURF for robust three dimensional classification. In Proc. ECCV, 2010. 2
  19. 19.J. J. Koenderink and A. J. Van Doorn. The singularities of the visual mapping. Biological cybernetics, 24(1):51–59, 1976. 3
  20. 20.A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In Proc. NIPS. 2012. 3, 4, 6
  21. 21.D. G. Lowe. Object recognition from local scale-invariant features. In Proc. ICCV, 1999. 1, 3
  22. 22.D. Macrini, A. Shokoufandeh, S. Dickinson, K. Siddiqi, and S. Zucker. View-based 3-D object recognition using shock graphs. In Proc. ICPR, volume 3, 2002. 3
  23. 23.H. Murase and S. K. Nayar. Visual learning and recognition of 3-D objects from appearance. 14(1), 1995. 3
  24. 24.R. Osada, T. Funkhouser, B. Chazelle, and D. Dobkin. Shape distributions. ACM Trans. Graph., 21, 2002. 2
  25. 25.F. Perronnin, J. Sanchez, and T. Mensink. Improving the ´ Fisher kernel for large-scale image classification. In Proc. ECCV, 2010. 1, 3
  26. 26.B. T. Phong. Illumination for computer generated pictures. Commun. ACM, 18(6), 1975. 3
  27. 27.A. S. Razavin, H. Azizpour, J. Sullivan, and S. Carlsson. Cnn features off-the-shelf: An astounding baseline for recognition. In DeepVision workshop, 2014. 3
  28. 28.J. Sanchez, F. Perronnin, T. Mensink, and J. Verbeek. Image classification with the fisher vector: Theory and practice. 2013. 4
  29. 29.R. G. Schneider and T. Tuytelaars. Sketch classification and classification-driven analysis using fisher vectors. ACM Trans. Graph., 33(6):174:1–174:9, Nov. 2014. 2, 3, 7, 8
  30. 30.T. Shao, W. Xu, K. Yin, J. Wang, K. Zhou, and B. Guo. Discriminative sketch-based 3D model retrieval via robust shape matching. In Computer Graphics Forum, volume 30. Wiley Online Library, 2011. 7
  31. 31.K. Simonyan, O. M. Parkhi, A. Vedaldi, and A. Zisserman. Fisher vector faces in the wild. In Proc. BMVC, 2013. 5
  32. 32.K. Simonyan, A. Vedaldi, and A. Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. CoRR, abs/1312.6034, 2013. 6
  33. 33.K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2014. 7
  34. 34.C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich. Going deeper with convolutions. CoRR, abs/1409.4842, 2014. 6
  35. 35.A. Vedaldi and B. Fulkerson. VLFeat: An open and portable library of computer vision algorithms. http://www.vlfeat.org/, 2008. 4
  36. 36.Z. Wu, S. Song, A. Khosla, F. Yu, L. Zhang, X. Tang, and J. Xiao. 3d shapenets: A deep representation for volumetric shape modeling. In Proc. CVPR, to appear, 2015. 1, 2, 3, 4, 5, 6, 9
  37. 37.S. M. Yoon, M. Scherer, T. Schreck, and A. Kuijper. Sketch-based 3D model retrieval using diffusion tensor fields of suggestive contours. In Proc. International Conference on Multimedia, 2010. 7

Citation

MLA
Su, H., et al. “Multi-view Convolutional Neural Networks for 3D Shape Recognition”. arXiv, 2015, http://arxiv.org/abs/1505.00880v3.
APA
Su, H., Maji, S., Kalogerakis, E., & Learned-Miller, E. (2015). Multi-view Convolutional Neural Networks for 3D Shape Recognition. arXiv. http://arxiv.org/abs/1505.00880v3
Chicago
Su, H., S. Maji, E. Kalogerakis, and E. Learned-Miller. 2015. “Multi-view Convolutional Neural Networks for 3D Shape Recognition”. arXiv. http://arxiv.org/abs/1505.00880v3.
Harvard
Su, H. et al. (2015) “Multi-view Convolutional Neural Networks for 3D Shape Recognition”, arXiv [Preprint]. Available at: http://arxiv.org/abs/1505.00880v3.
Vancouver
1. Su H, Maji S, Kalogerakis E, Learned-Miller E (2015) Multi-view Convolutional Neural Networks for 3D Shape Recognition. arXiv

BibTeX

@article{su2015multi,
  title = {Multi-view Convolutional Neural Networks for 3D Shape Recognition},
  author = {Su, Hang and Maji, Subhransu and Kalogerakis, Evangelos and Learned-Miller, Erik},
  year = {2015},
  journal = {arXiv},
  url = {http://arxiv.org/abs/1505.00880v3},
  eprint = {1505.00880}
}
Metadata:arXiv

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: IEEE