SPG: Unsupervised Domain Adaptation for 3D Object Detection via Semantic Point Generation

Xu, Qiangeng; Zhou, Yin; Wang, Weiyue; Qi, Charles R.; Anguelov, Dragomir

doi:10.48550/arxiv.2108.06709

Cited by 2 publications

(5 citation statements)

References 74 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…All of them are used to obtain new datasets through data processing methods and to improve the effectiveness of models. The semantic point generation (SPG) [42] method is used to generate semantic point cloud directly in the predicted region, and is used to reproduce the missing data from the original point cloud. Then the newly obtained semantic point cloud data and the original point cloud data are fused to obtain a new point cloud data.…”

Section: Data Processing Based Methodsmentioning

confidence: 99%

“…In T3D [37], VoteNet [38], CT3D [39], Group-Free [40], 3DETR [41] the architecture of Transformer was used for better feature extraction to overcome the difficulty of matching point cloud data and the difficulty of effective feature extraction. The SPG [42], rangeguided cylindrical network [43] processed the point cloud data to make up some defects of the original data to get a new dataset, which improved the final model effect from the perspective of the whole data. RangeDet [44], PV-RCNN [45], CenterPoint [46] enhance the detection method of feature maps from different perspectives.…”

Section: Other Methodsmentioning

confidence: 99%

“…In point cloud classification tasks, it provides important information for human-related tasks, such as point cloud action recognition P4Transformer [29], PST-Net [30], PoseNet [32], HandVoxNet [33], PVN3D [35], point cloud pose estimation PointContrast [24], HandVoxNet [33], PVN3D [35]. In point cloud detection and tracking, T3D [37], VoteNet [38], CT3D [39], Group-Free [40], 3DETR [41], SPG [42], range-guided cylindrical network [43], RangeDet [44], PV-RCNN [45], CenterPoint [46] can be applied to automatic driving, tracking, etc. In point cloud segmentation, Squeeze-Seg [47], geometric shared network (GS-Net) [48], VIASeg [49], LU-Net [50], RangeNet++ [51], JSNet [52], MPNet [53], SceneEncoder [54] perform the segmentation task well, these models perform more accurate extraction from global to local by performing feature extraction on point cloud data, main methods are adding attention mechanism, adding more information to the original point cloud data, and mining sequence information in spatial and temporal dimensions.…”

Section: Future Research Directionsmentioning

confidence: 99%

“…In HandVoxNet [33] researchers add depth maps to supplement the missing information of point cloud data. In SPG [42] semantic data and original point cloud data are fused for data tolerance to reproduce the original point cloud missing data. VIASeg [49] embedding colour information into the point cloud data provides rich visual information to the point cloud data.…”

Section: More Information For Point Cloud Datasetsmentioning

confidence: 99%

See 3 more Smart Citations

A survey on end‐to‐end point cloud learning

Tang

Huang

et al. 2022

IET Image Processing

View full text Add to dashboard Cite

Point cloud is an important expression form of three‐dimensional (3D) data. It has enjoyed continuous development and attracted increasing attention due to its wide applications in many areas, such as artificial intelligence, deep learning, autonomous driving and tracking. Recently, there is a large number of end‐to‐end point cloud‐based deep learning methods being proposed which are successful in the 3D domain. In order to better use point cloud data for analysis and to explore future research directions, this paper presents a comprehensive review of existing methods and publicly available datasets, with a focus on the methods and research status of using point cloud data as direct input. The background of point cloud is first introduced, including data acquisition methods, basic concepts, and challenges. Following that, the deep learning methods based on point cloud data are investigated and analysed according to classification, detection and tracking, and segmentation. Furthermore, the existing public datasets and evaluation metrics are introduced. Finally, promising research directions are proposed in conjunction with existing methods.

show abstract

Section: Data Processing Based Methodsmentioning

confidence: 99%

Section: Other Methodsmentioning

confidence: 99%

Section: Future Research Directionsmentioning

confidence: 99%

Section: More Information For Point Cloud Datasetsmentioning

confidence: 99%

See 2 more Smart Citations

A survey on end‐to‐end point cloud learning

Tang

Huang

et al. 2022

IET Image Processing

View full text Add to dashboard Cite

show abstract

“…Frustum based approaches improved direct Hdnet [355] 2018 Ca + Li 2 Stage 20 use of raw-point cloud on DNN however Frustum PointNets [245] 2018 Ca + Li Fusion 2.9 lacked processing speed for real-time Second [353] 2018 Ca + Li 2 Stage 40 embedded deployment & Applications Squeezeseg [338] 2018 Ca + Li Fusion 50 Pointpilllars [155] 2019 Ca + Li 1 Stage 25 (GTX 1080 Ti) PointRCNN [276] 2019 Ca + Li 1 Stage 10 Data Fusion pipelines improved the Lasernet [206] 2019 Ca + Li 1 Stage 83 segmentation application on point clouds Class-Balanced [404] 2019 Ca + Li 1 Stage 42 Sparse-to-dense [360] 2019 Ca + Li Fusion 10 Mono3d++ [97] 2019 Ca + Li 1 Stage 20 Approaches such as machine-learned pillar Pointpainting [319] 2020 Ca + Li 1 Stage 2.5 encoders are learned in an end-to-end manner SA-SSD [95] 2020 Ca + Li 1 Stage 25 Infofocus [323] 2020 Ca + Li 1 Stage 31 (GTX 1080 Ti) 3dSSD [358] 2020 Ca + Li 1 Stage 25 LiDAR 3d object detection networks heavily SE-SSD [396] 2021 Ca + Li 1 Stage 32 rely on labeled training data SPG [349] 2021 Ca + Li 1 Stage 41.56 Voxel-Transformer [203] 2021 Ca + Li 1 Stage 43 Pyramid-RCNN [201] 2021 Ca + Li 1 Stage -Grid based methods converts the point-Channel-wise [274] 2021 Ca + Li 1 Stage 39 cloud unstructured data to pixel & voxel Voxel-To-Point [167] 2021 Li 2 Stage 41 for 2D and 3D convolution processing Voxel-RCNN [56] 2021 Ca + Li 1 Stage 40.8 Multi-View to H-3D [57] 2021 Ca + Li 1 Stage Recent approach involves using encoders SA-Det3D [29] 2021 Ca + Li 1 Stage 36 for detection refinement of far and distant X-View [343] 2021 Ca + Li 1 Stage 47 objects, these decoders enhances the point CenterPoint [367] 2021 Ca + Li Fusion 16 feature through hierarchical aggregation.…”

Section: A Perceptionmentioning

confidence: 99%

A Survey on Approximate Edge AI for Energy Efficient Autonomous Driving Services

Katare

Perino

Nurmi

et al. 2023

IEEE Commun. Surv. Tutorials

View full text Add to dashboard Cite

Autonomous driving services depends on active sensing from modules such as camera, LiDAR, radar, and communication units. Traditionally, these modules process the sensed data on high-performance computing units inside the vehicle, which can deploy intelligent algorithms and AI models. The sensors mentioned above can produce large volumes of data, potentially reaching up to 20 Terabytes. This data size is influenced by factors such as the duration of driving, the data rate, and the sensor specifications. Consequently, this substantial amount of data can lead to significant power consumption on the vehicle. Similarly, a substantial amount of data will be exchanged between infrastructure sensors and vehicles for collaborative vehicle applications or fully connected autonomous vehicles. This communication process generates an additional surge of energy consumption. Although the autonomous vehicle domain has seen advancements in sensory technologies, wireless communication, computing and AI/ML algorithms, the challenge still exists in how to apply and integrate these technology innovations to achieve energy efficiency. This survey reviews and compares the connected vehicular applications, vehicular communications, approximation and Edge AI techniques. The focus is on energy efficiency by covering newly proposed approximation and enabling frameworks. To the best of our knowledge, this survey is the first to review the latest approximate Edge AI frameworks and publicly available datasets in energy-efficient autonomous driving. The insights from this survey can benefit the collaborative driving service development on low-power and memory-constrained systems and the energy optimization of autonomous vehicles.

show abstract

SPG: Unsupervised Domain Adaptation for 3D Object Detection via Semantic Point Generation

Abstract: Figure 1: Our Semantic Point Generation (SPG) recovers the foreground regions by generating semantic points (red). Combined with the original cloud, these semantic points can be directly used by modern LiDAR-based detectors and help improve the detection results (green boxes).

Cited by 2 publications

References 74 publications

A survey on end‐to‐end point cloud learning

A survey on end‐to‐end point cloud learning

A Survey on Approximate Edge AI for Energy Efficient Autonomous Driving Services

Contact Info

Product

Resources

About