Auto-Trajectory-Prediction-Devkit

This repository contains related algorithms in prediction fields:

• LaneGCN
• LaneGCN checkpoints
• Vectornet
• Vectornet checkpoints
• TNT
• TNT checkpoints
• mmtransformer
• mmtransformer checkpoints

Table of Contents

• Auto-Trajectory-Prediction-Devkit
  • Table of Contents
  • Methods Difference
  • How to Use
  • Requirements

Methods Difference

Method	ADE_1	FDE_1	MR_1	mADE	mFDE	mMR
LaneGCN	1.34	2.95	0.49	0.71	1.08	0.10
VectorNet	1.54	3.44	0.57	-	-	-

• LaneGCN
  • More lightweight data preprocessing(official need more 40G+ memory)
  • Optimized part of GCN attention module
  • Add a module to supervise predict angles
  • Easy to distribute training
• VectorNet
  • Use transformer instead of GNNs
  • Support to convert onnx, easy to deploy on edge devices
  • Easy to distribute training

How to Use

# for lanegcn:
$ cd ./data
$ python lanegcn_preprocess.py -i /Argoverse/train -m train
$ python lanegcn_preprocess.py -i /Argoverse/val -m val
$ python train_ddp.py -f ./config/Config_Lanegcn.yaml

# for vectornet:
$ python train_ddp.py -f ./config/Config_VectorNet.yaml

For details about how to configure related algorithms, see examples.

Requirements

• requirements.txt