The provided code evaluates R-CNN and FV-CNN descriptors on various texture and material datasets (DTD, FMD, KTH-TIPS2b, ALOT), as well as for other datasets: objects (PASCAL VOC 2007), scenes (MIT Indoor), and fine-grained (CUB 200-2011). The results of these experiments are described in Table 1 and 2 of ** Cimpoi15 ** and Tables 3, 4, 5, and 6 of ** Cimpoi15a. **

After downloading the code, make sure that the dependencies are resolved (see below).

You also have to download the datasets you want to evaluate on, and link to them or copy them under data folder, in the location of your repository. Download the models (VGG-M, VGG-VD and AlexNet) in data/models. It is slightly faster to download them manually from here: http://www.vlfeat.org/matconvnet/pretrained.

Once done, run the run_experiments.m file.

In run_experiments.m you could remove (or add) dataset names to the datasetList cell. Make sure you adjust the number of splits accordingly. The datasets are specified as {'dataset_name', <num_splits>} cells.

The code relies on vlfeat, and matconvnet, which should be downloaded and built before running the experiments. Run git submodule update -i in the repository download folder.

To build vlfeat, go to <deep-fbanks-dir>/vlfeat and run make; ensure you have MATLAB executable and mex in the path.

To build matconvnet, go to <deep-fbanks-dir>/matconvnet/matlab and run vl_compilenn; ensure you have CUDA installed, and nvcc in the path.

For LLC features (Table 3 in arxiv paper), please download the code from http://www.robots.ox.ac.uk/~vgg/software/enceval_toolkit and copy the following to the code folder (no subfolders!)

• enceval/enceval-toolkit/+featpipem/+lib/LLCEncode.m
• enceval/enceval-toolkit/+featpipem/+lib/LLCEncodeHelper.cpp
• enceval/enceval-toolkit/+featpipem/+lib/annkmeans.m

Create the corresponding dcnnllc encoder type (see the examples provided in run_experiments.m for BOVW, VLAD or FV).

The <dataset-name>_get_database.m files generate the imdb structure for each dataset. Make sure the datasets are copied or linked to manually in the data folder for this to work.

The datasets are stored in individual folders under data, in the current code folder, and experiment results are stored in data/exp01 folder, in the same location as the code. Alternatively, you could make data and experiments symbolic links pointing to convenient locations.

Please be aware that the descriptors are stored on disk (in cache folder, under data/exp01/<experiment-dir>), and may require large amounts of free space (especially FV-CNN features).

Describable Textures Dataset (DTD) is publicly available for download at: http://www.robots.ox.ac.uk/~vgg/data/dtd. You can also download the precomputed DeCAF features for DTD, the paper and evaluation results.

Our additional annotations for OpenSurfaces dataset are publicly available for download at: http://www.robots.ox.ac.uk/~vgg/data/wildtex

Code for CVPR14 paper (and Table 2 in arXiv paper): http://www.robots.ox.ac.uk/~vgg/data/dtd/download/desctex.tar.gz

If you use the code and data please cite the following in your work:

FV-CNN code and additional annotaitons for the OpenSurfaces dataset:

@article{Cimpoi15a,
Author       = "Cimpoi, M. and Maji, S., Kokkinos, I. and Vedaldi, A.",
Title        = "Deep Filter Banks for Texture Recognition, Description, and Segmentation"
Journal      = "arXiv preprint arXiv:1507.02620",
Year         = "2015",
}

@inproceedings{Cimpoi15,
Author       = "Cimpoi, M. and Maji, S. and Vedaldi, A.",
Title        = "Deep Filter Banks for Texture Recognition and Segmentation",
Booktitle    = "IEEE Conference on Computer Vision and Pattern Recognition",
Year         = "2015",
}

DTD dataset and IFV + DeCAF baseline:

@inproceedings{cimpoi14describing,
Author       = "M. Cimpoi and S. Maji and I. Kokkinos and S. Mohamed and A. Vedaldi",
Title        = "Describing Textures in the Wild",
Booktitle    = "IEEE Conference on Computer Vision and Pattern Recognition",
Year         = "2014",
}