FastPillars: A Deployment-friendly Pillar-based 3D Detector

Zhou, Sifan; Tian, Zhi; Chu, Xiangxiang; Zhang, Xinyu; Zhang, Bo; Lu, Xinzheng; Feng, Chengjian; Jie, Zequn; Chiang, Patrick; Ma, Lin

doi:10.48550/arxiv.2302.02367

Cited by 8 publications

(19 citation statements)

References 39 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Historically, pillar-based approaches trail voxel-based methods in terms of performance. Recently, [25], [39], [56] introduce more advanced backbones, bridging the performance gap with voxel-based methods.…”

Section: A Grid-based 3d Object Detectionmentioning

confidence: 99%

“…Voxel-based methods [9], [48], [53], [57] produce 3D voxels and employ 3D convolutions for feature extraction. Pillar-based [24], [25], [39], [56] approaches first transform the point clouds into a pseudo-image representation and then employ a 2D backbone for feature extraction.…”

Section: Introductionmentioning

confidence: 99%

“…2.7M PillarNet [39] 60.2% 17.1M PillarNeSt-Base 63.2% 8.9 M FastPillar [56] 61.7% 11.6M PillarNeSt-Large 64.3%…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

PillarNeSt: Embracing Backbone Scaling and Pretraining for Pillar-based 3D Object Detection

Mao,

Wang,

Zhang

et al. 2024

IEEE Trans. Intell. Veh.

View full text Add to dashboard Cite

Pillar-based 3D object detectors mainly employ randomly initialized 2D convolution neural network (ConvNet) for feature extraction and fail to enjoy the benefits from the backbone scaling and pretraining in the image domain. This paper shows the effectiveness of 2D backbone scaling and pretraining for pillar-based 3D object detectors. For better backbone scaling, we first introduce several design principles for point cloud backbone, to tackle the sparsity of point cloud and improve the effective receptive field. The backbone scaling is achieved by adaptively designed based on the model size. For backbone pretraining, we propose a weight adaptation module, to transfer the image knowledge obtained by pretraining on large-scale image datasets for the point cloud. Our proposed pillar-based detector, termed PillarNeSt, outperforms the existing 3D object detectors by a large margin on the nuScenes and Argoversev2 datasets. Code is released at https://github.com/WayneMao/PillarNeSt.

show abstract

Section: A Grid-based 3d Object Detectionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

PillarNeSt: Embracing Backbone Scaling and Pretraining for Pillar-based 3D Object Detection

Mao,

Wang,

Zhang

et al. 2024

IEEE Trans. Intell. Veh.

View full text Add to dashboard Cite

show abstract

“…As the first detector that utilizes the concept of segmenting 3D space in pillars for 3D object detection in autonomous driving applications, PointPillars 13 employs PointNet 14 FastPillars 21 is an end-to-end trainable neural network optimized for the deployment issues of 3D detectors in autonomous driving applications. It does not rely on sparse convolution and achieves a balance between detection accuracy and runtime efficiency through a redesigned robust architecture.…”

Section: Introductionmentioning

confidence: 99%

Pillar-based multilayer pseudo-image 3D object detection

Guo,

Lu,

Huang

et al. 2024

J. Electron. Imag.

View full text Add to dashboard Cite

Real-time and reliable three-dimensional (3D) object detection is critical for the autonomous driving system. Many 3D object detectors have recently adopted the pillar-based method represented by PointPillars due to its benefits of simplicity, speed, modularity, and ease of expansion. However, the detection effectiveness of the pillar-based method is limited by the lack of intra-pillar feature learning and insufficient multiscale pseudo-image feature extraction. Therefore, we present a multilayer pseudo-image 3D object detection method to enhance the performance of the pillar-based method. First, the proposed detector utilizes a multiscale feature extraction module to learn the point cloud representation within the pillars. The pillarized point cloud space is then converted into two pseudo-images of the same size, each encoding unique spatial information. After being individually processed via an inter-channel attention mechanism, these two pseudo-images get concatenated to form a global pseudo-feature map. Finally, a multiscale feature extraction backbone network with upper-and lower-level fusion is employed to process the global pseudo-feature map into a high-level representation. The multi-task detection head then generates the final detection results. Experimental findings on the KITTI dataset validate the effectiveness and superiority of the proposed method. The 3D mAP, BEV mAP, and mAOS of the proposed method are improved by 2.85%, 2.48%, and 2.76% compared to the traditional PointPillars, respectively.

show abstract

“…However, voxelization requires sampling and compression of point cloud, which may lose some detail information and requires more computational and storage space. To overcome these drawbacks, some pillar-based object detection methods have been proposed, [13][14][15][16] which convert point clouds into pillar grids and use CNN networks to extract features. However, pillar grids may lead to uneven sampling or grid bias, which may affect the accuracy of object detection.…”

Section: Introductionmentioning

confidence: 99%

Enhancing small object detection in point clouds with self-attention voting network

Zhu,

Wang,

et al. 2024

Opt. Eng.

View full text Add to dashboard Cite

The development of point cloud-based object detection in the field of autonomous driving has been rapid. However, it is undeniable that the issue of detecting small objects with high precision remains an urgent challenge. To address this issue, we introduce a single-stage 3D detection network, termed self-attention voting-single stage detection (SAV-SSD). It directly extracts feature information from the raw point cloud data and introduces an innovative self-attention voting mechanism to generate center points through weighted voting based on feature correlations. Compared with the feature prediction, we make an additional prediction of the center point, which can better control the position and size of the bounding boxes to improve the accuracy and stability of the predictions. To capture more features of small objects, cross multi-scale feature fusion is designed to establish connections between deep and shallow features. Experimental results demonstrate that SAV-SSD significantly improves the accuracy of pedestrian and cyclist detection while maintaining realtime performance. On the KITTI dataset, SAV-SSD outperforms many state-ofthe-art 3D object detection methods.

show abstract

FastPillars: A Deployment-friendly Pillar-based 3D Detector

Cited by 8 publications

References 39 publications

PillarNeSt: Embracing Backbone Scaling and Pretraining for Pillar-based 3D Object Detection

PillarNeSt: Embracing Backbone Scaling and Pretraining for Pillar-based 3D Object Detection

Pillar-based multilayer pseudo-image 3D object detection

Enhancing small object detection in point clouds with self-attention voting network

Contact Info

Product

Resources

About