Geometry-Aware Self-Training for Unsupervised Domain Adaptation on Object Point Clouds

Zou, Longkun; Tang, Hui; Chen, Ke; Jia, Kui

doi:10.1109/iccv48922.2021.00634

Cited by 41 publications

(24 citation statements)

References 23 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…PointDAN [43] has been the first work to address point cloud classification in the UDA context; they leverage on the well known Maximum Classifier Discrepancy (MCD) [45] to achieve alignment in feature space. Differently, [2,1,73] exploit Self-Supervised Learning (SSL) to run additional tasks on both domains. [4] also leverages on point cloud reconstruction, but uses it to refine pseudo-labels.…”

Section: Related Workmentioning

confidence: 99%

“…of models trained on synthetic CAD data and then tested on point clouds obtained with real sensors, where parts of the object may be missing due to occlusions and measurements are corrupted by noise. Here comes to help Unsupervised Domain Adaptation (UDA), which pursues solving a supervised learning task in a Target domain, T , where data come without labels, by leveraging on labeled data available in a Source domain, S. In the last couple of years, an increasing number of papers [4,43,1,2,73] have addressed UDA for point cloud classification, with popular synthetic datasets of CAD models, such as ModelNet40 [65] or ShapeNet [7], and real datasets such as ScanNet [11]. The main line of research focuses on learning an effective feature space for the target domain by means of auxiliary tasks such as point cloud reconstruction [1,46], 3D puzzle sorting [2] and rotation prediction [73].…”

Section: Introductionmentioning

confidence: 99%

“…Here comes to help Unsupervised Domain Adaptation (UDA), which pursues solving a supervised learning task in a Target domain, T , where data come without labels, by leveraging on labeled data available in a Source domain, S. In the last couple of years, an increasing number of papers [4,43,1,2,73] have addressed UDA for point cloud classification, with popular synthetic datasets of CAD models, such as ModelNet40 [65] or ShapeNet [7], and real datasets such as ScanNet [11]. The main line of research focuses on learning an effective feature space for the target domain by means of auxiliary tasks such as point cloud reconstruction [1,46], 3D puzzle sorting [2] and rotation prediction [73]. These tasks are refereed as auxiliary since they do not directly solve the main task, but at the same time, they are useful to learn features for the target domain without the need of annotations.…”

Section: Introductionmentioning

confidence: 99%

“…In the second step of our proposal, following recently published works in the field [4,73,46,17], we make use of self-training, an iterative methodology that exploits the predictions of a pre-trained model (pseudo-labels) to provide partial supervision on the target domain as well. However, pseudo-labels are noisy and their naive use typically leads to overfitting of the dominant classes of the source domain as shown in [71,75].…”

Section: Introductionmentioning

confidence: 99%

“…However, pseudo-labels are noisy and their naive use typically leads to overfitting of the dominant classes of the source domain as shown in [71,75]. The strategy proposed by [4] to refine them requires offline training of an additional network for this purpose and the definition of hand-crafted rules based on k-NN queries, limiting its general applicability, while [73] adopted a standard procedure borrowed from the 2D world [74]. As a further contribution of our work, we take a different path, and propose to use Graph Neural Networks (GNNs) [63] to refine pseudo-labels online during self-training.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Self-Distillation for Unsupervised 3D Domain Adaptation

Cardace¹,

Spezialetti²,

Ramirez³

et al. 2022

Preprint

View full text Add to dashboard Cite

Point cloud classification is a popular task in 3D vision. However, previous works, usually assume that point clouds at test time are obtained with the same procedure or sensor as those at training time. Unsupervised Domain Adaptation (UDA) instead, breaks this assumption and tries to solve the task on an unlabeled target domain, leveraging only on a supervised source domain. For point cloud classification, recent UDA methods try to align features across domains via auxiliary tasks such as point cloud reconstruction, which however do not optimize the discriminative power in the target domain in feature space. In contrast, in this work, we focus on obtaining a discriminative feature space for the target domain enforcing consistency between a point cloud and its augmented version. We then propose a novel iterative self-training methodology that exploits Graph Neural Networks in the UDA context to refine pseudo-labels. We perform extensive experiments and set the new state-of-the art in standard UDA benchmarks for point cloud classification. Finally, we show how our approach can be extended to more complex tasks such as part segmentation.

show abstract

Section: Related Workmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Self-Distillation for Unsupervised 3D Domain Adaptation

Cardace¹,

Spezialetti²,

Ramirez³

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

Learning Cross-Domain Features for Domain Generalization on Point Clouds

Xiao

Cheng

Shi

2022

Pattern Recognition and Computer Vision

View full text Add to dashboard Cite

CoSMix: Compositional Semantic Mix for Domain Adaptation in 3D LiDAR Segmentation

Saltori

Galasso

Fiameni³

et al. 2022

Lecture Notes in Computer Science

View full text Add to dashboard Cite

Deep-learning models for 3D point cloud semantic segmentation exhibit limited generalization capabilities when trained and tested on data captured with different sensors or in varying environments due to domain shift. Domain adaptation methods can be employed to mitigate this domain shift, for instance, by simulating sensor noise, developing domain-agnostic generators, or training point cloud completion networks. Often, these methods are tailored for range view maps or necessitate multi-modal input. In contrast, domain adaptation in the image domain can be executed through sample mixing, which emphasizes input data manipulation rather than employing distinct adaptation modules. In this study, we introduce compositional semantic mixing for point cloud domain adaptation, representing the first unsupervised domain adaptation technique for point cloud segmentation based on semantic and geometric sample mixing. We present a two-branch symmetric network architecture capable of concurrently processing point clouds from a source domain (e.g. synthetic) and point clouds from a target domain (e.g. real-world). Each branch operates within one domain by integrating selected data fragments from the other domain and utilizing semantic information derived from source labels and target (pseudo) labels. Additionally, our method can leverage a limited number of human point-level annotations (semi-supervised) to further enhance performance. We assess our approach in both synthetic-to-real and real-to-real scenarios using LiDAR datasets and demonstrate that it significantly outperforms state-of-the-art methods in both unsupervised and semi-supervised settings.

show abstract

Geometry-Aware Self-Training for Unsupervised Domain Adaptation on Object Point Clouds

Cited by 41 publications

References 23 publications

Self-Distillation for Unsupervised 3D Domain Adaptation

Self-Distillation for Unsupervised 3D Domain Adaptation

Learning Cross-Domain Features for Domain Generalization on Point Clouds

CoSMix: Compositional Semantic Mix for Domain Adaptation in 3D LiDAR Segmentation

Contact Info

Product

Resources

About