Category-Level Adversaries for Outdoor LiDAR Point Clouds Cross-Domain Semantic Segmentation

Yuan, Zhimin; Wen, Chenglu; Cheng, Ming; Su, Yongxuan; Liu, Weiquan; Yu, Shangshu; Wang, Cheng

doi:10.1109/tits.2022.3219853

Cited by 9 publications

(4 citation statements)

References 43 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Our method demonstrates a significant 9.2% improvement in performance over LiDomAug, which relies on ego-motion and multi-frame information, highlighting its robustness and effectiveness. Moreover, although our approach falls short by 5.9% compared to the latest DA technique, DCFNet [8], in the (N→K) scenario, it excels in the (K→N) scenario, outperforming DCFNet by 13.9%. The qualitative results, as shown in Fig.…”

Section: Comparison To State-of-the-art Da/dg Methodsmentioning

confidence: 77%

“…2) Experiments in the LiDomAug Setting: We benchmark the proposed method against various augmentation methods, including CutMix [44], Copy-Paste [45], Mix3D [37], Po-larMix [46], and domain adaptation methods like SWD [47], 3DGCA [25], Complete & Label [11], LiDAR-UDA [30], CLAN [42], FADA [41], DAST [43] and DCFNet [8]. We employ the experimental setting proposed by LiDomAug [12], using SemanticKITTI [5] and nuScenes [6] datasets.…”

Section: Comparison To State-of-the-art Da/dg Methodsmentioning

confidence: 99%

“…IV shows the performance of various data augmentation methods, domain adaptation methods, and our proposed method, all using MinkNet42 as the backbone. The results for CLAN [42], FADA [41], DAST [43], and DCFNet [8] are taken from the DCFNet publication. While our method shows slightly lower performance than LiDomAug in the (N→K) scenario, it excels in the (K→N) scenario, surpassing all other augmentation methods.…”

Section: Comparison To State-of-the-art Da/dg Methodsmentioning

confidence: 99%

“…Rochan et al [29] propose a range view-based unsupervised domain adaptation method, aligning beam positions between training and target data, thereby ensuring spatial consistency in sensor data for model adaptation. Yuan et al [8] present a multi-scale feature extraction technique to handle LiDAR density variations and propose an adversarial learning weight mechanism, capitalizing on consistency in class predictions across point and voxel representations. Shaban et al [30] introduce LiDAR-UDA, combining point cloud subsampling to mimic various LiDAR types and refining pseudo labels through Cross-frame Ensembling and a Learned Aggregation Model (LAM).…”

Section: B Lidar-based Domain Adaptationmentioning

confidence: 99%

See 3 more Smart Citations

Single Domain Generalization for LiDAR Semantic Segmentation

Kim,

Kang,

et al. 2023

2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

View full text Add to dashboard Cite

While significant progress has been achieved in LiDAR-based perception, domain generalization continues to present challenges, often resulting in reduced performance when encountering unfamiliar datasets due to domain discrepancies. One of the primary hurdles stems from the variability of LiDAR sensors, leading to inconsistencies in point cloud density distribution. Such inconsistencies can undermine the effectiveness of perception models. We address this challenge by introducing a new approach that acknowledges a fundamental characteristic of LiDAR: the variation in point density due to the distance from the LiDAR to the scene, and the number of beams relative to the field of view. Understanding this, we view each LiDAR's point cloud at various distances as having distinct density distributions, which can be consistent across different LiDAR models. With this insight, we propose the Density Discriminative Feature Embedding (DDFE) module, crafted to specifically extract features related to density while ensuring domain invariance across different LiDAR sensors. In addition, we introduce a straightforward but effective density augmentation technique, designed to broaden the density spectrum and enhance the capabilities of the DDFE. The proposed DDFE stands out as a versatile and lightweight domain generalization module. It can be seamlessly integrated into various 3D backbone networks, consistently outperforming existing state-of-the-art domain generalization approaches. We commit to releasing the source code publicly to foster community collaboration and advancement.

show abstract

Section: Comparison To State-of-the-art Da/dg Methodsmentioning

confidence: 77%