3D Object Detection for Autonomous Driving: A Survey

Qian, Rui; Lai, Xin Ji; Li, Xirong

doi:10.1016/j.patcog.2022.108796

Cited by 221 publications

(67 citation statements)

References 1 publication

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…PointPillars [13] takes advantage of a standard 2D convolutional detection pipeline which codes point clouds into a special partition of voxels (i.e., pillars) in the light of the tradeoff between efficiency and accuracy. However, the step of point clouds discretization in voxel-based detectors will degrade the fine-grained localization accuracy [17]. Some point-voxel-based methods [20,4,35,10] have been proposed, which take the advantage of localization accuracy of point-based detectors and computational efficiency of voxel-based detectors.…”

Section: Lidar-based 3d Object Detectionmentioning

confidence: 99%

“…The 3D Sparse Convolutions [31] with positional encoding can focus on voxels with point clouds thus reducing computation cost. The common voxel representation learning operators include mean operator, random sampling, and MLP operator [17]. Reconstructing the voxel features as a regression task is very difficult as the pre-training network needs to learn the distributions of each point.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Voxel-MAE: Masked Autoencoders for Pre-training Large-scale Point Clouds

Chen¹,

Xu²,

Zhao³

et al. 2022

Preprint

View full text Add to dashboard Cite

Mask-based pre-training has achieved great success for self-supervised learning in image, video, and language, without manually annotated supervision. However, it has not yet been studied about large-scale point clouds with redundant spatial information in autonomous driving. As the number of large-scale point clouds is huge, it is impossible to reconstruct the input point clouds. In this paper, we propose a mask voxel classification network for large-scale point clouds pre-training. Our key idea is to divide the point clouds into voxel representations and classify whether the voxel contains point clouds. This simple strategy makes the network to be voxel-aware of the object shape, thus improving the performance of downstream task, such as 3D object detection. Our Voxel-MAE with even a 90% masking ratio can still learn representative features for the high spatial redundancy of large-scale point clouds. We also validate the effectiveness of Voxel-MAE in unsupervised domain adaptative task, which proves the generalization ability of Voxel-MAE. Our Voxel-MAE proves that it is feasible to pre-train large-scale point clouds without data annotations to enhance the perception ability of the autonomous vehicle. Extensive experiments show great effectiveness of our pre-trained model with 3D object detectors (SEC-OND, CenterPoint, and PV-RCNN) on two popular datasets (KITTI, Waymo). Codes are publicly available at https: //github.com/chaytonmin/Voxel-MAE.

show abstract

Section: Lidar-based 3d Object Detectionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Voxel-MAE: Masked Autoencoders for Pre-training Large-scale Point Clouds

Chen¹,

Xu²,

Zhao³

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

“…We briefly introduce the 3D object detection approaches in this section, and refer reader to Qian et al [19] for more detailed description. ImVoteNet [16] proposes to use 2D image RGB, geometric coordinates, semantics, and pixel texture information to assist 3D point clouds object detection.…”

Section: D Object Detectionmentioning

confidence: 99%

Label-Guided Auxiliary Training Improves 3D Object Detector

Huang¹,

Liu²,

Zhu³

et al. 2022

Preprint

View full text Add to dashboard Cite

Detecting 3D objects from point clouds is a practical yet challenging task that has attracted increasing attention recently. In this paper, we propose a Label-Guided auxiliary training method for 3D object detection (LG3D), which serves as an auxiliary network to enhance the feature learning of existing 3D object detectors. Specifically, we propose two novel modules: a Label-Annotation-Inducer that maps annotations and point clouds in bounding boxes to task-specific representations and a Label-Knowledge-Mapper that assists the original features to obtain detection-critical representations. The proposed auxiliary network is discarded in inference and thus has no extra computational cost at test time. We conduct extensive experiments on both indoor and outdoor datasets to verify the effectiveness of our approach. For example, our proposed LG3D improves VoteNet by 2.5% and 3.1% mAP on the SUN RGB-D and ScanNetV2 datasets, respectively. The code is available at https://github.com/FabienCode/LG3D.

show abstract

“…There exist many deep-learning-based object detection methods, and they have been applied in many fields, such as autonomous driving [ 3 ], facial recognition [ 4 ], defect detection [ 5 , 6 ], and medical imaging [ 7 , 8 ]; the historical champion network ResNet from the ImageNet LSVRC (Large-Scale Visual Recognition Challenge) competition outperformed the human recognition level [ 9 ]. Models based on CNNs have shown good accuracy and robustness for recognition of certain species.…”

Section: Introductionmentioning

confidence: 99%

Contour-Based Wild Animal Instance Segmentation Using a Few-Shot Detector

Tang

Zhao

et al. 2022

Animals

View full text Add to dashboard Cite

Camera traps are widely used in wildlife research, conservation, and management, and abundant images are acquired every day. Efficient real-time instance segmentation networks can help ecologists label and study wild animals. However, existing deep convolutional neural networks require a large number of annotations and labels, which makes them unsuitable for small datasets. In this paper, we propose a two-stage method for the instance segmentation of wildlife, including object detection and contour approximation. In the object detection stage, we use FSOD (few-shot object detection) to recognize animal species and detect the initial bounding boxes of animals. In the case of a small wildlife dataset, this method may improve the generalization ability of the wild animal species recognition and even identify new species that only have a small number of training samples. In the second stage, deep snake is used as the contour approximation model for the instance segmentation of wild mammals. The initial bounding boxes generated in the first stage are input to deep snake to approximate the contours of the animal bodies. The model fuses the advantages of detecting new species and real-time instance segmentation. The experimental results show that the proposed method is more suitable for wild animal instance segmentation, in comparison with pixel-wise segmentation methods. In particular, the proposed method shows a better performance when facing challenging images.

show abstract

3D Object Detection for Autonomous Driving: A Survey

Cited by 221 publications

References 1 publication

Voxel-MAE: Masked Autoencoders for Pre-training Large-scale Point Clouds

Voxel-MAE: Masked Autoencoders for Pre-training Large-scale Point Clouds

Label-Guided Auxiliary Training Improves 3D Object Detector

Contour-Based Wild Animal Instance Segmentation Using a Few-Shot Detector

Contact Info

Product

Resources

About