Tensorflow Implementation For "SGCN: Sparse Graph Convolution Network for Pedestrian Trajectory Prediction"

Project for Vision and Perception, DIAG, Sapienza University in Roma

Table of Contents

• Introduction
• Running
• Result
• Team

Introduction

This project is the tensorflow implementation for paper "SGCN:Sparse Graph Convolution Network for Pedestrian Trajectory Prediction" in CVPR 2021, and we also use a new dataset MOT-15 to test it.

Here is the original pytorch code. We rewrite the original author's code by tensorflow.

Paper

Running

Keep this github file structure and run the file "Main_SGCN_MOT15.ipynb" directly by jupyter notebook.

"metrics.py" implements loss functions.

"model.py" implements network model.

"utils.py" processes dataset.

"Visualization.ipynb" shows the visualization of the trojectory.

"dataset/data" file includes MOT-15 dataset，which has been processed.

"dataset/ETH" file includes eth dataset.

"dataset/hotel" file includes hotel dataset.

Result

Warning: due to using the different DL frame, tensorflow version is not exactly the same with pytorch version. Like the way to intial convolution layer kernel.

ADE & FDE

Metric\Dataset	ETH	HOTEL	MOT-15
ADE	0.83	0.49	0.14
FDE	1.56	0.78	0.25

Visulization

Green line is the target trojectory and the red line is the prediction trojectory. The result where the two trajectories seem to overlap is good.

Team

• PK
• SCC