SigFormer: Sparse Signal-Guided Transformer for MultiModal Human Action Segmentation

November 29, 2023 · View on GitHub

Introduction

This is an implementation repository for our work. SigFormer: Sparse Signal-Guided Transformer for MultiModal Human Action Segmentation.

Clone the repository and move to folder:

git clone https://github.com/LIUQI-creat/SigFormer.git

cd SigFormer

To use this source code, you need Python3.8+ and a few python3 packages:

Please download the OpenPack dataset use:

optk-download -d ./data

Use the following commands for training:

python src/train.py

Obtain the final prediction results:

python src/ensemble_mean.py

In order to get the results in the table below, you need to submit the generated submission.zip file to the online review.

Our submitted file is provided in baiduyun, passcode:ubfo.

F1 (Macro Average)	U0104	U0108	U0110	U0203	U0204	U0207	ALL
SigFormer	0.971	0.969	0.960	0.966	0.903	0.923	0.958