PointAcc: Efficient Point Cloud Accelerator

Lin, Yujun; Zhang, Zhekai; Tang, Haotian; Wang, Hanrui; Han, Song

doi:10.1145/3466752.3480084

Cited by 51 publications

(11 citation statements)

References 47 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Deep Learning for Point Clouds Point cloud algorithms are increasingly moving toward DNNs, which has spurred recent interests in accelerating point cloud networks [18,29,36]. Point cloud DNNs mainly come in two forms: one that operates on raw points [37,48,49,59,70], and the other that first voxelizes points and operates on voxels, which are grid-aligned points [16,22].…”

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

Crescent: Taming Memory Irregularities for Accelerating Deep Point Cloud Analytics

Ye¹,

Hammonds²,

Gan³

et al. 2022

Preprint

View full text Add to dashboard Cite

3D perception in point clouds is transforming the perception ability of future intelligent machines. Point cloud algorithms, however, are plagued by irregular memory accesses, leading to massive inefficiencies in the memory sub-system, which bottlenecks the overall efficiency.This paper proposes Crescent, an algorithm-hardware co-design system that tames the irregularities in deep point cloud analytics while achieving high accuracy. To that end, we introduce two approximation techniques, approximate neighbor search and selectively bank conflict elision, that "regularize" the DRAM and SRAM memory accesses. Doing so, however, necessarily introduces accuracy loss, which we mitigate by a new network training procedure that integrates approximation into the network training process. In essence, our training procedure trains models that are conditioned upon a specific approximate setting and, thus, retain a high accuracy. Experiments show that Crescent doubles the performance and halves the energy consumption compared to an optimized baseline accelerator with < 1% accuracy loss.

show abstract

Section: Related Workmentioning

confidence: 99%

“…PointAcc [36], Point-X [69], and Mesorasi [18] are all recent point cloud accelerators. They are fundamentally orthogonal to our work in that they focus on accelerating the feature computation in point cloud DNNs.…”

Section: Related Workmentioning

confidence: 99%

Crescent: Taming Memory Irregularities for Accelerating Deep Point Cloud Analytics

Ye¹,

Hammonds²,

Gan³

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

“…A method based on combining the architectures of volumetric and multi-view neural networks is presented in the work [28]. Volumetric neural networks are also actively used to work with three-dimensional point clouds [29,30]. In [31] the task of searching for a query object of unknown position and pose in a scene, both given in the form of 3D point cloud data, was studied.…”

Section: Volumetric Neural Networkmentioning

confidence: 99%

Semantic 3D Reconstruction of a Scene and Its Effective Visualisation

Kniaz

Moshkantsev

Bordodymov

et al. 2022

Proceedings of the 32nd International Conference on Computer Graphics and Vision

View full text Add to dashboard Cite

Single-image 3D scene reconstruction is required in multiple challenging tasks including mobile robotics, industrial monitoring and reconstruction of lost cultural heritage. While modern models demonstrate robust resolution of scene in real time with resolution up to 128 x 128 x 128 voxels, visualization of such detailed of a such detailed voxel model is challenging. A model with 128<sup>3</sup> voxels contains 2097152 simple cubes 16M vertices. It is unfeasible for modern hardware to perform visualization of such voxel models in real-time. Hence a voxel model simplification technique is required to demonstrate reconstruction results in real-time. In this paper, we propose a new algorithm for voxel model simplification using predefined camera views. The algorithm reduces a rigid-body voxel model to a shell voxel model. It keeps only the voxels that are visible from the required view. We demonstrate the effectiveness of the proposed algorithm using a case study with a mobile robot and a state-of-the-art SSZ single-photo 3D reconstruction neural network. We generated a real and a virtual scene with various objects including a statue. We use a mobile robot equipped with a single camera to collect real and synthetic data. We train the SSZ model using the collected data. We developed a dedicated visualization software that implements our algorithm. The comparison of the visualization performance for the full model and its reduced version demonstrates that our algorithm allows to increase the performance by 420 times.

show abstract

“…The above works are based on the PointNet and PointNet++ networks, and thus cannot be directly applied to the acceleration of SSCN. PointAcc [20] proposed an ASIC-based accelerator that unified diverse mapping operations into a multiply-accumulate operation through coordinate transformation to be compatible with different point cloud networks. Other hardware solutions such as GPUs can be deployed to accelerate the point cloud networks.…”

Section: Introductionmentioning

confidence: 99%

“…However, because the restricted computation pattern of the Sub-Conv layer leads to irregular sparse matching operations, traditional convolution accelerators suffer from performance degradation when they are directly applied to it [20]. Therefore, efficient accelerators for SSCN are urgently needed, and the bottleneck lies in the extreme and unstructured sparsity, and the complex computational dependency between the sparsity of the central activation and the neighborhood ones.…”

Section: Introductionmentioning

confidence: 99%

An Efficient FPGA Accelerator for Point Cloud

Mao

Yang

Wang

et al. 2022

2022 IEEE 35th International System-on-Chip Conference (SOCC)

View full text Add to dashboard Cite

Deep learning-based point cloud processing plays an important role in various vision tasks, such as autonomous driving, virtual reality (VR), and augmented reality (AR). The submanifold sparse convolutional network (SSCN) has been widely used for the point cloud due to its unique advantages in terms of visual results. However, existing convolutional neural network accelerators suffer from non-trivial performance degradation when employed to accelerate SSCN because of the extreme and unstructured sparsity, and the complex computational dependency between the sparsity of the central activation and the neighborhood ones. In this paper, we propose a high performance FPGA-based accelerator for SSCN. Firstly, we develop a zero removing strategy to remove the coarse-grained redundant regions, thus significantly improving computational efficiency. Secondly, we propose a concise encoding scheme to obtain the matching information for efficient point-wise multiplications. Thirdly, we develop a sparse data matching unit and a computing core based on the proposed encoding scheme, which can convert the irregular sparse operations into regular multiplyaccumulate operations. Finally, an efficient hardware architecture for the submanifold sparse convolutional layer is developed and implemented on the Xilinx ZCU102 field-programmable gate array board, where the 3D submanifold sparse U-Net is taken as the benchmark. The experimental results demonstrate that our design drastically improves computational efficiency, and can dramatically improve the power efficiency by 51 times compared to GPU.

show abstract

PointAcc: Efficient Point Cloud Accelerator

Cited by 51 publications

References 47 publications

Crescent: Taming Memory Irregularities for Accelerating Deep Point Cloud Analytics

Crescent: Taming Memory Irregularities for Accelerating Deep Point Cloud Analytics

Semantic 3D Reconstruction of a Scene and Its Effective Visualisation

An Efficient FPGA Accelerator for Point Cloud

Contact Info

Product

Resources

About