# VoxelNeXt: Fully Sparse VoxelNet for 3D Object Detection and Tracking

**Authors:** Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, Jiaya Jia  
**arXiv:** [2303.11301](https://arxiv.org/abs/2303.11301)  
**Date:** 2023-03-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.11301v1/

## Key points

- Traditional 3D object detectors rely on hand-crafted proxies and dense prediction heads, leading to increased computational costs.
- VoxelNext is a fully sparse 3D object detection framework that directly predicts objects based on sparse voxel features without the need for hand-crafted proxies.
- The core of VoxelNext is the utilization of a strong sparse convolutional network called VoxelNeXt for detecting and tracking 3D objects solely through voxel features.
- VoxelNext eliminates the need for sparse-to-dense conversion or post-processing steps like Non-Maximum Suppression (NMS), resulting in an elegant and efficient detection framework.
- VoxelNext achieves a superior speed-accuracy trade-off compared to other mainstream detectors on the nuScenes dataset.
- The study demonstrates the effectiveness of fully sparse voxel-based representation for LIDAR 3D object detection and tracking, marking a significant advancement in this field.
- Extensive experiments validate the efficacy of VoxelNext on nuScenes, Waymo, and Argoverse2 benchmarks, outperforming all existing LIDAR methods on the nuScenes tracking test benchmark.
- The availability of code and models facilitates reproducibility and adoption of this innovative approach within the research community.

## Summary

In their paper titled "VoxelNeXt: Fully Sparse VoxelNet for 3D Object Detection and Tracking," authors Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia address the limitations of traditional 3D object detectors that rely on hand-crafted proxies and dense prediction heads. These methods require converting sparse voxel features to denser representations, leading to increased computational costs. To overcome these challenges, the authors propose VoxelNext - a fully sparse 3D object detection framework that directly predicts objects based on sparse voxel features without the need for hand-crafted proxies. The core insight of VoxelNext lies in its utilization of a strong sparse convolutional network called VoxelNeXt which enables the detection and tracking of 3D objects solely through voxel features. This approach eliminates the necessity for sparse-to-dense conversion or post-processing steps such as Non-Maximum Suppression (NMS), resulting in an elegant and efficient detection framework. The authors demonstrate that VoxelNext achieves a superior speed-accuracy trade-off compared to other mainstream detectors on the nuScenes dataset. Furthermore, the study showcases the effectiveness of a fully sparse voxel-based representation for LIDAR 3D object detection and tracking - marking a significant advancement in this field. Extensive experiments conducted on nuScenes, Waymo, and Argoverse2 benchmarks validate the efficacy of the proposed approach. Remarkably without any additional embellishments - the model outperforms all existing LIDAR methods on the nuScenes tracking test benchmark. Overall,VoxelNeXt presents a novel and efficient solution for 3D object detection by leveraging sparse voxel features directly. The findings highlight the potential of fully sparse representations in enhancing performance while reducing computational complexity in this domain. The availability of code and models further facilitates reproducibility and adoption of this innovative approach within the research community.