Unsupervised Domain Adaptation in LiDAR Semantic Segmentation with Self-Supervision and Gated Adapters

Rochan, Mrigank; Aich, Shubhra; Corral-Soto, Eduardo R.; Nabatchian, Amir; Liu, Bingbing

doi:10.1109/icra46639.2022.9811654

Cited by 12 publications

(16 citation statements)

References 21 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Since there is a lack of extensive research on real-to-real domain adaptation in 3D LiDAR point clouds, there is no standard baseline with a predefined set of object classes. Every method tried to create its new object categories by merging pre-existing object labels; Therefore, we compare our method with two stateof-the-art UDA methods, namely UDASSGA [20], com-plete&Label (C&L) [31] and also with other methods such as (M+A)Ent [28], SWD [16] and CORAL [22] that were reported in the recent UDA paper UDASSGA [20]. To make the comparison fair, we have divided the experiments into two sections according to the used subsets of classes on which the state-of-the-art methods are evaluated, see IV-C and IV-D.…”

Section: Methodsmentioning

confidence: 99%

“…Unsupervised Domain Adaptation (UDA) for Point Cloud Segmentation. UDA for point cloud segmentation can be classified into two setups: (1) simulation-to-real [21], [33] and (2) real-to-real [31], [28], [20]. Simulation-to-real UDA is used when a deep learning-based model is trained with source domain from simulated or synthetically generated data and then tested on a target domain real-world data.…”

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

Teachers in Concordance for Pseudo-Labeling of 3D Sequential Data

Gebrehiwot

Vacek

Hurych

et al. 2023

IEEE Robot. Autom. Lett.

View full text Add to dashboard Cite

Automatic pseudo-labeling is a powerful tool to tap into large amounts of sequential unlabeled data. It is especially appealing in safety-critical applications of autonomous driving, where performance requirements are extreme, datasets are large, and manual labeling is very challenging. We propose to leverage sequences of point clouds to boost the pseudolabeling technique in a teacher-student setup via training multiple teachers, each with access to different temporal information. This set of teachers, dubbed Concordance, provides higher quality pseudo-labels for student training than standard methods. The output of multiple teachers is combined via a novel pseudolabel confidence-guided criterion. Our experimental evaluation focuses on the 3D point cloud domain and urban driving scenarios. We show the performance of our method applied to 3D semantic segmentation and 3D object detection on three benchmark datasets. Our approach, which uses only 20% manual labels, outperforms some fully supervised methods. A notable performance boost is achieved for classes rarely appearing in training data. Our codes will be made publicly available.

show abstract

Section: Methodsmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Teachers in Concordance for Pseudo-Labeling of 3D Sequential Data

Gebrehiwot

Vacek

Hurych

et al. 2023

IEEE Robot. Autom. Lett.

View full text Add to dashboard Cite

show abstract

“…After projection, RV images can be processed with existing 2D convolution networks. RV-based networks are affected by domain shift, which can be mitigated by using generative approaches [19], [24], feature alignment [19], [20], [24], and contrastive learning [27]. RayCast [24] tackles the real-toreal UDA problem by transferring the sensor pattern of the target domain to the source domain through ray casting.…”

Section: Domain Adaptation For Point Cloud Segmentationmentioning

confidence: 99%

“…SqueezeSegV2 [20] improves the SqueezeSeg [43] architecture, and reduces domain shift in the synth-to-real setup by aligning source and target features with a geodesic RayCast [24] real-to-real RV RangeNet++ [25] Sem.KITTI [5] nuSc. [7] ✓ ✓ ✓ ePointDA [19] synth-to-real RV SqueezeSegV2 [20] GTA-V [20] KITTI [26] ✓ ✓ ✓ ✓ Sem.KITTI [5] SqueezeSegV2 [20] synth-to-real RV SqueezeSegV2 [20] GTA-V [20] KITTI [26] ✓ ✓ ✓ Gated [27] synth-to-real RV SalsaNext [28] GTA-V [20] KITTI [26] ✓ ✓ ✓ real-to-real nuScenes [7] KITTI [26] nuScenes [7] xMUDA [29] real-to-real 2D&3D xMUDA [29] nuSc. [7] nuSc.…”

Section: Domain Adaptation For Point Cloud Segmentationmentioning

confidence: 99%

“…At feature level, a higher-order momentum loss [63] is used to learn domain agnostic features between source and target input data. Gated [27] states that domain shift between source and target point clouds can be mitigated by solving the sparsity shift and by introducing domain specific parameters. Given an input pair of source an target RV images, they first solve the sparsity difference thorough selfsupervised completion and by applying a dropout mask.…”

Section: Domain Adaptation For Point Cloud Segmentationmentioning

confidence: 99%

See 1 more Smart Citation

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