Tramac / Non-local-tensorflow

Non-local Block for Semantic Segmentation

Non-local-tensorflow

An implementation of Non-local block intruduced in Tensorflow. Link to the original paper:Non-local Neural Networks

Introduction

This repository includes the code(training and testing) for DRN and Non-local. The code is for semantic segmentation.

Requirements

python 2.7.x
tensorflow >= 1.4.0

Usage

1. Download VOC2012 dataset.
   

2. Prepare the tfrecord data to train the model. Please see my another repository Convert_To_TFRecord.
   

3. Train the model
   

#the parameter of --mode in main.py need to be "train"
python main.py

4. Test the model
   

#the parameter of --mode in main.py need to be "test"
python main.py

Experiments

Method	Non-local	mIoU
DRN	✗	7.31
DRN	✓	0.094

The result is so terrible, non-local has a bad influence, I have no idea about it, what's wrong with this? Is it because there is no pretrainig?

To Do

• COCO
• Cityscapes
• ADE

Reference

• Non-local_pytorch