Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution

By Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, Song Han. European Conference on Computer Vision, 2020

863 citations recorded by the source corpus.

Self-driving cars need to understand 3D scenes efficiently and accurately in order to drive safely. Given the limited hardware resources, existing 3D perception models are not able to recognize small instances (e.g., pedestrians, cyclists) very well due to the low-resolution voxelization and aggressive downsampling. To this end, we propose Sparse Point-Voxel Convolution (SPVConv), a lightweight 3D module that equips the vanilla Sparse Convolution with the high-resolution point-based branch. With negligible overhead, this point-based branch is able to preserve the fine details even from large outdoor scenes. To explore the spectrum of efficient 3D models, we first define a flexible architecture design space based on SPVConv, and we then present 3D Neural Architecture Search (3D-NAS) to search the optimal network architecture over this diverse design space efficiently and effectively. Experimental results validate that the resulting SPVNAS model is fast and accurate: it outperforms the state-of-the-art MinkowskiNet by 3.3%, ranking 1st on the competitive SemanticKITTI leaderboard. It also achieves 8x computation reduction and 3x measured speedup over MinkowskiNet with higher accuracy. Finally, we transfer our method to 3D object detection, and it achieves consistent improvements over the one-stage detection baseline on KITTI.

Subfields: Aerospace Engineering, Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/769dbcb80cc1d2d17ae7c524644103d0f6595452

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Aerospace Engineering

1. Hongsheng Li
2. Song Han
3. Zhijian Liu
4. Yujun Lin
5. Xinge Zhu
6. Tai Wang
7. Dahua Lin
8. Jie Zhou

In Computer Vision and Pattern Recognition

1. Song Han
2. Luc Van Gool
3. Hongsheng Li
4. Yi Yang
5. Jie Zhou
6. Jiwen Lu
7. Xiaogang Wang
8. Dacheng Tao

Citations in the corpus

• NeuralRecon: Real-Time Coherent 3D Reconstruction from Monocular Video (2021, methodology): “We use torchsparse [43] as the implementation of 3D sparse convolution.”
• QuantumNAS: Noise-Adaptive Search for Robust Quantum Circuits (2021, methodology): “SuperCircuitbased search is inspired by the supernet method in classical ML model training [46]–[51].”
• BEVFusion: Multi-Task Multi-Sensor Fusion with Unified Bird's-Eye View Representation (2022, methodology): “There are also U-Net like models specialized for 3D semantic segmentation [17, 13, 52, 33, 75], an important task for offline HD map construction.”
• Tri-Perspective View for Vision-Based 3D Semantic Occupancy Prediction (2023, background): “Voxel representation [26,43,48] describes a 3D scene with dense cubic features V ∈ H × W × D × C where H , W , D are the spatial resolution of the voxel space and C denotes the feature dimension.”
• Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR Segmentation (2020, methodology): “…to some voxel partition and 3D convolution based methods, including FusionNet [48], TORANDONet [13] (multi-view fusion based method) and SPVNAS [33] (utilizing the neural architecture search for LiDAR segmentation), the proposed method also performs better than these 3D convolution based…”
• Sparse Single Sweep LiDAR Point Cloud Segmentation via Learning Contextual Shape Priors from Scene Completion (2020, methodology): “Also (Tang et al. 2020) use NAS to obtain a more efficient feature representation.”
• 2DPASS: 2D Priors Assisted Semantic Segmentation on LiDAR Point Clouds (2022, methodology): “[10] combines point-wise MLPs in each sparse convolution block to learn a point-voxel representation and uses NAS to search for a more efficient architecture.”
• RPVNet: A Deep and Efficient Range-Point-Voxel Fusion Network for LiDAR Point Cloud Segmentation (2021, methodology): “More recently, some efforts [10, 30] have been made to accelerate the 3D convolution, and improve the performance to a higher level with less computation.”
• SurroundOcc: Multi-Camera 3D Occupancy Prediction for Autonomous Driving (2023, methodology): “It has empowered the success of numerous methods on the lidar segmentation [33, 52, 11, 62, 61] and 3D scene completion [7, 46, 9, 25, 59, 29, 28] tasks.”