Awesome

NEW related work!!

Cross-Spectral Local Descriptors via Quadruplet Network

Learning cross-spectral similarity measures with deep convolutional neural networks

PDF | webpage

Bibtex

@inproceedings{Aguilera_cvprw_2016,
    organization = { IEEE  },
    year = { 2016  },
    pages = { 9  },
    month = { Jun  },
    location = { Las vegas, USA  },
    booktitle = { The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops  },
    author = { Cristhian A. Aguilera and Francisco J. Aguilera and Angel D. Sappa and Cristhian Aguilera and Ricardo Toledo  },
    title = { Learning cross-spectral similarity measures with deep convolutional neural networks  },
}

Instructions

First install the torch framework and cudnn

1. Install torch
2. Cudnn torch

Datasets

Nirscenes patches

Follow the following steps to generate the dataset

1. Download gt csv

   cd datasets
   ./download_nirscenes_csv.sh
   

2. Download the original nirscenes dataset link

3. Decompress the dataset on /datasets/nirscenes

4. Convert the images in each folder to ppm (Torch doesn't support tiff). Use your favorite software

5. Install csvigo

   luarocks install csvigo
   

6. Use our script to generate the dataset

   cd utils
   th nirscenes_to_t7.lua
   

The t7 dataset is stored in datasets/nirscenes

VIS-LWIR ICIP2015

1. Download the dataset

cd datasets
./download_icip_dataset.sh

It's easier if you are the owner of the dataset :)

Eval

Nirscenes eval (cpu and cuda support)

Evaluation code can be found in the eval folder. To eval one sequence:

1. You have to generate the nirscenes patch dataset

2. Install xlua

   luarocks install xlua
   

3. Run

   cd eval
   th nirscenes_eval.lua -seq_path ../dataset/nirscenes/[sequence].t7 -net .. [trained network]
   

For example, to eval the field sequence using the 2ch_country network bash th nirscenes_eval.lua -seq_path ../dataset/nirscenes/field.t7 -net ../trained_networks/2ch_country.net -net_type 2ch

For more options, run bash th nirscenes_eval -h

VIS-LWIR eval (ICIP2015) (just cuda support)

1. You have to download the dataset first

2. Run

   cd eval
   th icip2015_eval.lua -dataset_path ../nirscenes/icip2015/ -net [trained network]
   

For example. To eval 2ch_country

```bash
cd eval
th icip2015_eval.lua -dataset_path ../nirscenes/icip2015/ -net ../trained_networks/2ch_country.t7
```

Training

1. Install penlight

   luarocks install penlight
   

2. Train a network

   cd train
   th nirscenes_doall.lua -training_sequences [country|field|...] -net [2ch|siam|psiam]
   

For example, train a 2ch network using the country sequence

cd train
th nirscenes_doall.lua -training_sequences country -net 2ch

Results will be stored in the results folder.For more options, run

th nirscenes_doall.lua -h

Note The training code is different from the one used in the article. This new version is much faster.