RobustPointSet: A Dataset for Benchmarking Robustness of Point Cloud Classifiers

Taghanaki, Saeid Asgari; Luo, Jieliang; Zhang, Ran; Wang, Ye; Jayaraman, Pradeep Kumar; Jatavallabhula, Krishna Murthy

doi:10.48550/arxiv.2011.11572

Cited by 4 publications

(7 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In Partnet, we evaluate on Chair, Table , and Lamp because of high number of samples for these classes. Finally, we use the recently released RobustPointSet [45] dataset for evaluating transferablity of our network. Shapes in this dataset have missing regions synthetically created on various objects.…”

Section: Experiments and Resultsmentioning

confidence: 99%

See 1 more Smart Citation

Multimodal Shape Completion via IMLE

Arora,

Mishra,

Peng

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

Section: Experiments and Resultsmentioning

confidence: 99%

“…Transfer Learning. To examine whether our model can be used on other datasets, possibly with more non-uniform missing parts and noises, we perform an experiment on Robustpointset dataset [45]. Robustpointset contains various artifacts including missing regions, rotations, noise, etc.…”

Section: Experiments and Resultsmentioning

confidence: 99%

Multimodal Shape Completion via IMLE

Arora,

Mishra,

Peng

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…CT also achieves state-of-the-art performance in domain generalization on the VLCS dataset [Torralba and Efros, 2011]. We further test CT on corrupted 3D point-cloud data [Taghanaki et al, 2020] and show it outperforms existing methods in terms of mean classification accuracy over multiple test sets. To the best of our knowledge, this is the first work to explore both theoretically and empirically why BN leads to a model's over-reliance on brittle, low-variance features, which can adversely affect its performance on the downstream classification task.…”

Section: Introductionmentioning

confidence: 93%

“…CT's classification performance on corrupted 3D point cloud data. In this experiment, we use the RobustPointSet dataset [Taghanaki et al, 2020] which is created for analysis of point classifiers in terms of robustness to 3D corruptions. We follow the same training-domain validation setting as in [Taghanaki et al, 2020].…”

Section: Regularizing a Batch Normalized Model Using Ct Leads To A Hi...mentioning

confidence: 99%

“…Our results mark a significant improvement over prior works, which have tackled similar problems by either: (a) modifying batch normalization statistics of a trained model [Schneider et al, 2020, Benz et al, 2021 using privileged information from the target domain; or (b) augmenting the training data using a set of predefined corruption functions . As recent studies [Vasiljevic et al, 2016, Geirhos et al, 2018, Taghanaki et al, 2020 show that such approaches often fail to generalize due to the tendency of neural networks to memorize data-specific properties, this motivates a shift towards developing models that are inherently robust, regardless of particular data augmentations or input transformations applied during training.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Counterbalancing Teacher: Regularizing Batch Normalized Models for Robustness

Taghanaki¹,

Gholami²,

Khani³

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

Batch normalization (BN) is a ubiquitous technique for training deep neural networks that accelerates their convergence to reach higher accuracy. However, we demonstrate that BN comes with a fundamental drawback: it incentivizes the model to rely on low-variance features that are highly specific to the training (in-domain) data, hurting generalization performance on outof-domain examples. In this work, we investigate this phenomenon by first showing that removing BN layers across a wide range of architectures leads to lower out-of-domain and corruption errors at the cost of higher in-domain errors. We then propose Counterbalancing Teacher (CT), a method which leverages a frozen copy of the same model without BN as a teacher to enforce the student network's learning of robust representations by substantially adapting its weights through a consistency loss function. This regularization signal helps CT perform well in unforeseen data shifts, even without information from the target domain as in prior works. We theoretically show in an overparameterized linear regression setting why normalization leads to a model's reliance on such in-domain features, and empirically demonstrate the efficacy of CT by outperforming several baselines on robustness benchmarks such as CIFAR-10-C, CIFAR-100-C, and VLCS.

show abstract

To Drop or to Select: Reduce the Negative Effects of Disturbance Features for Point Cloud Classification From an Interpretable Perspective

Zhang

et al. 2023

IEEE Access

View full text Add to dashboard Cite

The perturbation features limit the performance of point cloud classification models both for clean point cloud and point cloud obtained from real-world. In this paper, we propose two methods to enhance models by reducing the negative impact of the nuisance features from the perspective of interpretability, namely, dropping the nuisance points before inputting the point cloud into models and adaptively selecting the important features during the training process. The former is achieved by saliency analysis of models, the perturbation points are those with low contribution to models. For each sample, dropping the low contribution points based on the saliency scores is equivalent to filtering the perturbation features. We design a generic framework for generating saliency maps for various models and datasets, and obtain empirical values for the number of dropped points on each set of them. Then we apply the unsupervised dropping process to improve the robustness of models. The latter is achieved by adaptive downsampling, and we design a multi-stage learnable class-attention-based downsampling module to replace the commonly used Farthest Point Sampling (FPS). As the training progresses, the downsampling module tends to select the common features for each category, thus eliminating the nuisance features to improve the learning efficiency of the model. For dropping points (DP), we generate saliency maps for PointNet&++, DGCNN and PointMLP on ModelNet40 and ScanObjectNN, PointNet+DP reaches an overall accuracy (OA) of 92.5% and 72% on ModelNet40 and ScanObjectNN, surpassing the original model by 3.4% and 5.3%, the OA of PointNet++ with DP on ModelNet40 object classification can be raised from 91.8% to 93.7%. For adaptive feature selection (AFS), PointMLP-elite+AFS reaches an OA of 92.5% and a mean accuracy (mAcc) of 72% on ScanObjectNN, surpassing the original model by 0.8% and 1%. It reaches the level of PointMLP with 6.3% of the number of parameters. Considering the difficulty of the deployment of deep models, PointMLP-elite+AFS is the most cost-effective classification model known on ScanObjectNN. INDEX TERMS Dropping points, Adaptive feature selection, Point cloud denoising, Point cloud classification VOLUME 4, 2016

show abstract

RobustPointSet: A Dataset for Benchmarking Robustness of Point Cloud Classifiers

Cited by 4 publications

References 0 publications

Multimodal Shape Completion via IMLE

Multimodal Shape Completion via IMLE

Counterbalancing Teacher: Regularizing Batch Normalized Models for Robustness

To Drop or to Select: Reduce the Negative Effects of Disturbance Features for Point Cloud Classification From an Interpretable Perspective

Contact Info

Product

Resources

About