Home

Awesome

ScatterFormer: Efficient Voxel Transformer with Scattered Linear Attention

This repo is the official implementation of paper: ScatterFormer: Efficient Voxel Transformer with Scattered Linear Attention. It achieves 73.8 mAP L2 on Waymo Open Dataset val and 72.4 NDS on NuScenes val. The scatterformer achieve real-time speed of 23 FPS.

<!-- > ScatterFormer: Efficient Voxel Transformer with Scattered Linear Attention > > [Chenhang He*](https://skyhehe123.github.io/), Ruihuang Li, Guowen Zhang, Lei Zhang --> <div align="center"> <img src="assets/performance.png" width="400"/> </div>

News

Main results

Waymo Open Dataset validation

Model#SweepsmAP/H_L1mAP/H_L2Veh_L1Veh_L2Ped_L1Ped_L2Cyc_L1Cyc_L2Log
ScatterFormer (100%)181.8/79.775.7/73.881.0/80.573.1/72.784.5/79.977.0/72.679.9/78.977.1/76.1Log
ScatterFormer (20%)180.3/78.074.1/72.079.6/79.171.6/71.283.5/78.375.9/71.077.7/76.674.8/73.7Log

NuScenes validation

ModelmAPNDSmATEmASEmAOEmAVEmAAEckptLog
ScatterFormer68.372.426.524.524.723.318.8ckptLog

Usage

Installation

Please refer to INSTALL.md for installation.

Dataset Preparation

Please follow the instructions from OpenPCDet. We adopt the same data generation process.

Sparse Group-wise Convolution

ScatterFormer relies on a group-wise sparse convolution, please find this hacked version of spconv

Training

# multi-gpu training
cd tools
bash scripts/dist_train.sh 8 --cfg_file <CONFIG_FILE> [other optional arguments]

Testing

# multi-gpu testing
cd tools
bash scripts/dist_test.sh 8 --cfg_file <CONFIG_FILE> --ckpt <CHECKPOINT_FILE>