This repository contains all the necessary to reproduce the work "Implicit neural representation for change detection" by P. Naylor, D. Di Carlo, A. Traviglia, M. Yamada and M. Fiorucci. You can find the paper here.
Most of the required packages are listed in the environment.yml file, and Nextflow should be installed to launch the pipeline successfully.
Please set the nextflow.config file to match your environment. We use two profiles, one for our local environment (debugging and prototyping) and a second for the cluster when we launch extensive experiments.
These files can be found in the exp_config folder and are YAML files.
The first parameters correspond to the config file, the output folder, the data path and the format.
The following parameters control the training and hyper-parameters.
In particular, a hyperparameter is a categorical list or a range given by a min and a max.
For each Nextflow file, a process will be spawned for each single_method, double_method and feature_method.
The pipeline accepts txt files as input in the following format: NAME_0.txt and NAME_1.txt.
Modify the variable paired_txt in the file nextflow/main.nf.
Please set the extension parameter to txt.
The pipeline accepts ply files as input in the following format: NAME_0.ply and NAME_1.ply.
Please set the extension parameter to ply.
We make available two experiments:
The data Urb3DCD can be accessed here. We used version 1 and 2 for the paper.
In particular, it is Urb3DCD-v2 for the colab.
To reproduce the tables presented in the paper and the plots, we first have to run for each configuration and dataset:
make paper_home
This command will produce a CSV file result/benchmark.csv with a line per method, dataset, and associated metrics.
To produce the boxplots shown in Figure 4 and Table 1 and A.1, please run the following:
python python/plots/boxplots.py --csv benchmark.csv
In addition, in the folder result/paper, you should find distribution plots and maps corresponding to Figures 3 and 5.
To run the ablation study presented in Section 5.2, please run the following:
make paper_ablation
To produce the benchmark_ablation.csv file, which contains the metrics for each dataset, configuration and lambda values.
Finally, to produce Figure 6, please run the following:
python python/plots/hyper_parameter.py --csv benchmark_ablation.csv
If you use our code, please cite us!
@inproceedings{naylor2024implicit,
title={Implicit neural representation for change detection},
author={Naylor, Peter and Di Carlo, Diego and Traviglia, Arianna and Yamada, Makoto and Fiorucci, Marco},
booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
pages={935--945},
year={2024}
}