Fully Convolutional Siamese Network for Scene Change Detection

Requirements

• Python2.7
• Pytorch0.2.0_3 (see: pytorch installation instuctions)
• torchvision

Datasets

This repo is built for scene change detection. We report the performance on three datasets.

• PCD2015 dataset

  • paper: Change detection from a street image pair using cnn features and superpixel segmentation
  • dataset: http://www.vision.is.tohoku.ac.jp/us/research/4d_city_modeling/pano_cd_dataset/

• VL_CMU_CD dataset

  • paper: Street-view change detection with deconvolutional networks
  • dataset: https://ghsi.github.io/proj/RSS2016.html

• CD2014 dataset

  • paper: changedetection.net: A new change detection benchmark dataset
  • dataset: http://changedetection.net/

    06/12/2018 update

    We have uploaded the modified CD2014 dataset to [baiduyun][googledrive], if you find cd2014 dataset is useful for your research, please cite the paper:

  @inproceedings{Goyette2012changedetection, title={changedetection.net: A New Change Detection Benchmark Dataset}, author={Goyette, Nil and Jodoin, Pierre Marc and Porikli, Fatih and Konrad, Janusz and Ishwar, Prakash}, booktitle={Computer Vision and Pattern Recognition Workshops}, pages={1-8}, year={2012}, }

Directory Structure

File Structure is as follows:

$T0_image_path/*.jpg
$T1_image_path/*.jpg
$ground_truth_path/*.jpg

Pretrained Model

Backbone model, which is deeplabv2 [baiduyun] [googledriver]in our work, is available, you should download it and put it to /pretrain

Pretrained models for PCD2015 and VL_CMU_CD also have been available.

• PCD2015: [baiduyun] [googledrive]
• VL_CMU_CD: [baiduyun] [googledrive]

Training

cd $SCD_ROOT
python train.py