pith. sign in

arxiv: 1907.06371 · v1 · pith:3D523PFAnew · submitted 2019-07-15 · 💻 cs.CV

Mitigating the Hubness Problem for Zero-Shot Learning of 3D Objects

Ali Cheraghian , Shafin Rahman , Dylan Campbell , Lars Petersson This is my paper

Pith reviewed 2026-05-24 21:38 UTC · model grok-4.3

classification 💻 cs.CV
keywords zero-shot learninghubness problem3D point cloudsobject recognitiongeneralized zero-shot learningModelNetpoint cloud classification3D vision
0
0 comments X

The pith

A dedicated loss reduces the hubness problem in zero-shot learning for 3D point cloud objects.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper seeks to establish that zero-shot learning applied to 3D objects faces a more severe hubness problem than in 2D because 3D lacks equivalent large pre-training datasets, resulting in lower quality features that bias predictions toward few labels. To fix this, the authors introduce a loss function that specifically targets the hubness bias. This matters because 3D sensors increasingly capture objects in the wild that the system has never seen during training. If the loss works, it improves recognition of those unseen objects in both standard zero-shot and generalized zero-shot settings. Evaluations on ModelNet40, ModelNet10, McGill, and SHREC2015 show new state-of-the-art performance.

Core claim

The hubness problem, where a model is biased to predict only a few particular labels for most of the test instances, is even more severe for 3D recognition than for 2D recognition. One reason is that in 2D one can use pre-trained networks trained on large datasets like ImageNet, which produces high-quality features. However, in the 3D case there are no such large-scale, labelled datasets available for pre-training which means that the extracted 3D features are of poorer quality which, in turn, exacerbates the hubness problem. The authors therefore propose a loss to specifically address the hubness problem. Their method is effective for both Zero-Shot and Generalized Zero-Shot Learning, and a

What carries the argument

A loss term introduced to specifically address and mitigate the hubness problem in the prediction of 3D object labels.

If this is right

  • The proposed loss improves performance in zero-shot learning of 3D objects.
  • It also improves generalized zero-shot learning where both seen and unseen classes are tested.
  • New state-of-the-art results are achieved on ModelNet40, ModelNet10, McGill and SHREC2015.
  • Hubness bias is reduced without needing to change the feature extractor.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Future 3D feature extractors trained on larger data might reduce the need for such a corrective loss.
  • The approach could extend to other 3D tasks like segmentation if hubness appears there.
  • Real-world 3D scans with noise might test whether the loss still holds its effectiveness.

Load-bearing premise

The hubness problem stems primarily from poorer 3D features due to lack of large pre-training datasets and can be sufficiently corrected by adding one loss term without changing the feature extractor or introducing new biases.

What would settle it

If experiments show that the proposed loss does not reduce the hubness metric (such as the skewness of the label prediction distribution) or fails to increase accuracy on unseen classes in the tested 3D datasets, the central claim would be falsified.

Figures

Figures reproduced from arXiv: 1907.06371 by Ali Cheraghian, Dylan Campbell, Lars Petersson, Shafin Rahman.

Figure 4
Figure 4. Figure 4: Qualitative results of successful (green) and failed (red) cases of our proposed ZSL method. Implementation details1 : We used the following set-up during training in all of the ex￾periments. We used the Adam optimizer [14] with an initial learning rate of 0.001 and a batch size of 64. For the point cloud network, we used PointNet [27] with five shared mlp layers (64,64,64,128,1024) followed by a max pooli… view at source ↗
read the original abstract

The development of advanced 3D sensors has enabled many objects to be captured in the wild at a large scale, and a 3D object recognition system may therefore encounter many objects for which the system has received no training. Zero-Shot Learning (ZSL) approaches can assist such systems in recognizing previously unseen objects. Applying ZSL to 3D point cloud objects is an emerging topic in the area of 3D vision, however, a significant problem that ZSL often suffers from is the so-called hubness problem, which is when a model is biased to predict only a few particular labels for most of the test instances. We observe that this hubness problem is even more severe for 3D recognition than for 2D recognition. One reason for this is that in 2D one can use pre-trained networks trained on large datasets like ImageNet, which produces high-quality features. However, in the 3D case there are no such large-scale, labelled datasets available for pre-training which means that the extracted 3D features are of poorer quality which, in turn, exacerbates the hubness problem. In this paper, we therefore propose a loss to specifically address the hubness problem. Our proposed method is effective for both Zero-Shot and Generalized Zero-Shot Learning, and we perform extensive evaluations on the challenging datasets ModelNet40, ModelNet10, McGill and SHREC2015. A new state-of-the-art result for both zero-shot tasks in the 3D case is established.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 4 minor

Summary. The paper claims that the hubness problem is more severe in 3D zero-shot learning (ZSL) and generalized ZSL than in 2D due to poorer feature quality from the absence of large-scale pre-training datasets like ImageNet; it proposes a dedicated loss term to mitigate hubness, shows effectiveness on both ZSL and GZSL protocols, and reports new state-of-the-art results on ModelNet40, ModelNet10, McGill, and SHREC2015.

Significance. If the reported results hold, the contribution is a lightweight, empirically effective correction for hubness in 3D ZSL that does not require changing the feature extractor. The manuscript's internal consistency—standard splits, hubness quantification via established metrics, and separate ZSL/GZSL evaluations—supports the central claim without circular derivations or unstated assumptions that undermine the evaluation.

minor comments (4)
  1. [§3] §3 (method): the proposed loss is described at a high level; adding the explicit formula (including any weighting hyperparameter) would improve reproducibility.
  2. [Tables 1-2] Table 1 and Table 2: report the number of runs or standard deviations for the accuracy and hubness metrics to allow readers to assess stability of the SOTA claims.
  3. [§5.3] §5.3 (ablation): the text states the loss addresses hubness specifically, but an additional row showing performance when the loss is replaced by a generic regularizer would strengthen that attribution.
  4. [Figure 3] Figure 3: the t-SNE visualizations are useful but the caption should explicitly state the perplexity and whether the same random seed was used across methods.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive assessment of our work and the recommendation for minor revision. We are pleased that the evaluation protocols, hubness quantification, and separation of ZSL/GZSL results were viewed as internally consistent and supportive of the central claims.

Circularity Check

0 steps flagged

No circularity; empirical loss proposed and evaluated on public benchmarks

full rationale

The paper presents an empirical method consisting of a dedicated loss term to address hubness in 3D ZSL/GZSL, built on top of existing feature extractors and evaluated using standard protocols on public datasets (ModelNet40, ModelNet10, McGill, SHREC2015). No equations, derivations, or parameter-fitting steps are described that reduce by construction to the paper's own inputs, predictions, or self-citations. The central claim remains independent of any self-definitional or fitted-input circularity, with hubness quantified via standard external metrics and results reported as SOTA under conventional splits.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities; the central claim rests on the unstated assumption that a single loss term can correct feature-induced hubness.

pith-pipeline@v0.9.0 · 5817 in / 1062 out tokens · 24152 ms · 2026-05-24T21:38:41.767165+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 45 canonical work pages · 5 internal anchors

  1. [1]

    Tensorflow: a system for large- scale machine learning

    Martín Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, et al. Tensorflow: a system for large- scale machine learning. In OSDI, volume 16, pages 265–283, 2016

    work page 2016

  2. [2]

    Evaluation of output embeddings for fine-grained image classification

    Zeynep Akata, Scott Reed, Daniel Walter, Honglak Lee, and Bernt Schiele. Evaluation of output embeddings for fine-grained image classification. InCVPR, pages 2927–2936, June 2015

    work page 2015

  3. [3]

    Label-embedding for image classification

    Zeynep Akata, Florent Perronnin, Zaid Harchaoui, and Cordelia Schmid. Label-embedding for image classification. TPAMI, 38(7):1425–1438, July 2016

    work page 2016

  4. [4]

    Synthesized classifiers for zero- shot learning

    Soravit Changpinyo, Wei-Lun Chao, Boqing Gong, and Fei Sha. Synthesized classifiers for zero- shot learning. In CVPR, pages 5327–5336, June 2016

    work page 2016

  5. [5]

    An empirical study and analysis of generalized zero-shot learning for object recognition in the wild

    Wei-Lun Chao, Boqing Changpinyo, Soravitand Gong, and Fei Sha. An empirical study and analysis of generalized zero-shot learning for object recognition in the wild. In ECCV, 2016

    work page 2016

  6. [6]

    Calibrate multiple consumer rgb-d cameras for low-cost and efficient 3d indoor mapping.Remote Sensing, 10(2), 2018

    Chi Chen, Bisheng Yang, Shuang Song, Mao Tian, Jianping Li, Wenxia Dai, and Lina Fang. Calibrate multiple consumer rgb-d cameras for low-cost and efficient 3d indoor mapping.Remote Sensing, 10(2), 2018

    work page 2018

  7. [7]

    3dcapsule: Extending the capsule architecture to classify 3d point clouds

    Ali Cheraghian and Lars Petersson. 3dcapsule: Extending the capsule architecture to classify 3d point clouds. In WACV, pages 1194–1202, Jan 2019

    work page 2019

  8. [8]

    Zero-shot learning of 3d point cloud objects

    Ali Cheraghian, Shafin Rahman, and Lars Petersson. Zero-shot learning of 3d point cloud objects. In MVA, 2019

    work page 2019

  9. [9]

    Attributes2classname: A discriminative model for attribute-based unsupervised zero-shot learning

    Berkan Demirel, Ramazan Gokberk Cinbis, and Nazli Ikizler-Cinbis. Attributes2classname: A discriminative model for attribute-based unsupervised zero-shot learning. In ICCV, Oct 2017

    work page 2017

  10. [10]

    Zero shot learning via multi-scale manifold regularization

    Shay Deutsch, Soheil Kolouri, Kyungnam Kim, Yuri Owechko, and Stefano Soatto. Zero shot learning via multi-scale manifold regularization. In CVPR, July 2017

    work page 2017

  11. [11]

    Improving zero-shot learning by mitigating the hubness problem

    Georgiana Dinu and Marco Baroni. Improving zero-shot learning by mitigating the hubness problem. in ICLR workshop, 2014

    work page 2014

  12. [12]

    Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

    Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  13. [13]

    Kinectfusion: Real-time 3d reconstruction and interaction using a moving depth camera

    Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, and Andrew Fitzgibbon. Kinectfusion: Real-time 3d reconstruction and interaction using a moving depth camera. In A. CHERAGHIAN ET. AL: MITIGA TING THE HUBNESS PROBLEM 11 Proceedings of the 24th Annual AC...

    work page 2011

  14. [14]

    Adam: A Method for Stochastic Optimization

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  15. [15]

    Semantic autoencoder for zero-shot learning

    Elyor Kodirov, Tao Xiang, and Shaogang Gong. Semantic autoencoder for zero-shot learning. In CVPR, July 2017

    work page 2017

  16. [16]

    Lampert, Hannes Nickisch, and Stefan Harmeling

    Christoph H. Lampert, Hannes Nickisch, and Stefan Harmeling. Learning to detect unseen object classes by between-class attribute transfer. In CVPR Workshops, pages 951–958, 2009

    work page 2009

  17. [17]

    Lampert, Hannes Nickisch, and Stefan Harmeling

    Christoph H. Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. TPAMI, pages 453–465, March 2014

    work page 2014

  18. [18]

    Multi-label zero-shot learning with structured knowledge graphs

    Chung-Wei Lee, Wei Fang, Chih-Kuan Yeh, and Yu-Chiang Frank Wang. Multi-label zero-shot learning with structured knowledge graphs. In CVPR, June 2018

    work page 2018

  19. [19]

    Chen, and Gim Hee Lee

    Jiaxin Li, Ben M. Chen, and Gim Hee Lee. So-net: Self-organizing network for point cloud analysis. In CVPR, pages 9397–9406, 2018

    work page 2018

  20. [20]

    Zero-shot recognition using dual visual-semantic mapping paths

    Yanan Li, Donghui Wang, Huanhang Hu, Yuetan Lin, and Yueting Zhuang. Zero-shot recognition using dual visual-semantic mapping paths. In CVPR, July 2017

    work page 2017

  21. [21]

    Z. Lian, J. Zhang, S. Choi, H. ElNaghy, J. El-Sana, T. Furuya, A. Giachetti, R. A. Guler, L. Lai, C. Li, H. Li, F. A. Limberger, R. Martin, R. U. Nakanishi, A. P. Neto, L. G. Nonato, R. Ohbuchi, K. Pevzner, D. Pickup, P. Rosin, A. Sharf, L. Sun, X. Sun, S. Tari, G. Unal, and R. C. Wilson. Non-rigid 3D Shape Retrieval. In I. Pratikakis, M. Spagnuolo, T. Th...

    work page 2015

  22. [22]

    Improving semantic embedding consistency by metric learning for zero-shot classification

    Stephane Herbin Maxime Bucher and Frederic Jurie. Improving semantic embedding consistency by metric learning for zero-shot classification. In ECCV, 2016

    work page 2016

  23. [23]

    Distributed represen- tations of words and phrases and their compositionality

    Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed represen- tations of words and phrases and their compositionality. In NIPS, pages 3111–3119. 2013

    work page 2013

  24. [24]

    Hinton, and Tom M

    Mark Palatucci, Dean Pomerleau, Geoffrey E. Hinton, and Tom M. Mitchell. Zero-shot learning with semantic output codes. In NIPS, pages 1410–1418, 2009

    work page 2009

  25. [25]

    Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glove: Global vectors for word representation. In EMNLP, pages 1532–1543, 2014

    work page 2014

  26. [26]

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

    Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NIPS, pages 5099–5108, 2017

    work page 2017

  27. [27]

    Charles Ruizhongtai Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017

    work page 2017

  28. [28]

    Hubs in space: Popular nearest neighbors in high-dimensional data

    Milos Radovanovic, Alexandros Nanopoulos, and Mirjana Ivanovic. Hubs in space: Popular nearest neighbors in high-dimensional data. JMLR, pages 2487–2531, 2010

    work page 2010

  29. [29]

    Zero-shot object detection: Learning to simul- taneously recognize and localize novel concepts

    Shafin Rahman, Salman Khan, and Fatih Porikli. Zero-shot object detection: Learning to simul- taneously recognize and localize novel concepts. In ACCV, December 2018

    work page 2018

  30. [30]

    A unified approach for conventional zero-shot, generalized zero-shot, and few-shot learning

    Shafin Rahman, Salman Khan, and Fatih Porikli. A unified approach for conventional zero-shot, generalized zero-shot, and few-shot learning. TIP, pages 5652–5667, Nov 2018. 12 A. CHERAGHIAN ET. AL: MITIGA TING THE HUBNESS PROBLEM

    work page 2018

  31. [31]

    Deep0tag: Deep multiple instance learning for zero-shot image tagging

    Shafin Rahman, Salman Khan, and Nick Barnes. Deep0tag: Deep multiple instance learning for zero-shot image tagging. IEEE Transactions on Multimedia, 2019

    work page 2019

  32. [32]

    An embarrassingly simple approach to zero-shot learning

    Bernardino Romera-Paredes and PHS Torr. An embarrassingly simple approach to zero-shot learning. In ICML, pages 2152–2161, 2015

    work page 2015

  33. [33]

    Ridge re- gression, hubness, and zero-shot learning

    Yutaro Shigeto, Ikumi Suzuki, Kazuo Hara, Masashi Shimbo, and Yuji Matsumoto. Ridge re- gression, hubness, and zero-shot learning. In ECMLPKDD, pages 135–151. Springer, 2015

    work page 2015

  34. [34]

    Retrieving articulated 3-d models using medial surfaces

    Kaleem Siddiqi, Juan Zhang, Diego Macrini, Ali Shokoufandeh, Sylvain Bouix, and Sven Dick- inson. Retrieving articulated 3-d models using medial surfaces. MVA, pages 261–275, May 2008

    work page 2008

  35. [35]

    Accelerating t-sne using tree-based algorithms

    Laurens Van Der Maaten. Accelerating t-sne using tree-based algorithms. JMLR, 15(1):3221– 3245, 2014

    work page 2014

  36. [36]

    The Caltech- UCSD Birds-200-2011 Dataset

    Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The Caltech- UCSD Birds-200-2011 Dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011

    work page 2011

  37. [37]

    Local Spectral Graph Convolution for Point Set Feature Learning

    Chu Wang, Babak Samari, and Kaleem Siddiqi. Local spectral graph convolution for point set feature learning. arXiv preprint arXiv:1803.05827, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  38. [38]

    Dynamic Graph CNN for Learning on Point Clouds

    Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph cnn for learning on point clouds. arXiv preprint arXiv:1801.07829 , 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  39. [39]

    Zhirong Wu, S. Song, A. Khosla, Fisher Yu, Linguang Zhang, Xiaoou Tang, and J. Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, pages 1912–1920, 2015

    work page 1912

  40. [40]

    Latent embeddings for zero-shot classification

    Yongqin Xian, Zeynep Akata, Gaurav Sharma, Quynh Nguyen, Matthias Hein, and Bernt Schiele. Latent embeddings for zero-shot classification. In CVPR, June 2016

    work page 2016

  41. [41]

    Lampert, Bernt Schiele, and Zeynep Akata

    Yongqin Xian, Christoph H. Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning - a comprehensive evaluation of the good, the bad and the ugly. TPAMI, 2018

    work page 2018

  42. [42]

    Monte carlo cross validation

    Qing-Song Xu and Yi-Zeng Liang. Monte carlo cross validation. Chemometrics and Intelligent Laboratory Systems, 56(1):1 – 11, 2001

    work page 2001

  43. [43]

    SpiderCNN: Deep Learning on Point Sets with Parameterized Convolutional Filters

    Yifan Xu, Tianqi Fan, Mingye Xu, Long Zeng, and Yu Qiao. Spidercnn: Deep learning on point sets with parameterized convolutional filters.arXiv preprint arXiv:1803.11527, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  44. [44]

    Learning a deep embedding model for zero-shot learning

    Li Zhang, Tao Xiang, and Shaogang Gong. Learning a deep embedding model for zero-shot learning. In CVPR, July 2017

    work page 2017

  45. [45]

    Domain- invariant projection learning for zero-shot recognition

    An Zhao, Mingyu Ding, Jiechao Guan, Zhiwu Lu, Tao Xiang, and Ji-Rong Wen. Domain- invariant projection learning for zero-shot recognition. In NIPS, 2018

    work page 2018