Spatial-Wise Dynamic Distillation for MLP-Like Efficient Visual Fault Detection of Freight Trains

This is the official implementation of Spatial-Wise Dynamic Distillation for MLP-Like Efficient Visual Fault Detection of Freight Trains.

Table of Contents

• Introduction
• Installation
• Usage
• Citation

---

Introduction

Fig. 2. Overview of the proposed MLP-like spatial-wise dynamic distillation method. We adopt a novel dynamic teacher architecture comprising three modules: label encoder, appearance encoder, and feature adaptive interaction. The dynamic teacher enables joint teacher-student training, which generates instructional representations from ground truth annotations and feature pyramids during the training stage.

Installation

This codebase is built upon Detectron2.

• Ubuntu 18.04 LTS, CUDA>=10.1, GCC>=7.5.0
• Python>=3.7.13
• Detectron2==0.3
• Pytorch>=1.7.1, torchvision>=0.8.2

1、Clone this repo：

git clone https://github.com/MVME-HBUT/SDD-FTI-FDet.git
cd SDD-FTI-FDet

2、Please install Detectron2 following the official guide: INSTALL.md.

3、Then create a conda virtual environment and activate it:

conda create -n SDD python=3.7 -y
conda activate SDD

4、Other requirements

pip3 install -r requirements.txt

Usage

Prepare your own datasets

For instance, downloading MS-COCO whose hierarchy is organized as follows:

MSCOCO
  |_ annotations
    |_ instances_train2017.json
    |_ instances_val2017.json
  |_ train2017
  |_ val2017

mkdir ${PROJ}/datasets
ln -s /path/to/MSCOCO datasets/coco

Train

python3 train.py --config-file configs/Distillation/SDD.yaml --num-gpus 1

Evaluation

It is handy to add [--eval-only] option to turn training command into evaluation usage.

python3 train.py --eval-only --config-file configs/Distillation/SDD.yaml --num-gpus 1

Visualization

python3 SDD-FTI-FDet/utils/result_vis.py

Citation

If you find this repository useful in your research, please consider citing:

@ARTICLE{10391271,
  author={Zhang, Yang and Pan, Huilin and Li, Mingying and Wang, An and Zhou, Yang and Ren, Hongliang},
  journal={IEEE Transactions on Industrial Electronics}, 
  title={Spatial-Wise Dynamic Distillation for MLP-Like Efficient Visual Fault Detection of Freight Trains}, 
  year={2024},
  volume={71},
  number={10},
  pages={13168-13177},
  keywords={Fault detection;Detectors;Feature extraction;Training;Visualization;Task analysis;Computational modeling;Dynamic distillation;fault detection;freight train images;multilayer perceptron (MLP)},
  doi={10.1109/TIE.2023.3344837}}