by Francesco Setti, Chris Russell, Chiara Bassetti and Marco Cristani

An F-formation arises whenever two or more people sustain a spatial and orientational relationship in which the space between them is one to which they have equal, direct, and exclusive access. [A. Kendon, 1990]

GCFF is a method to detect F-formations in static images, leveraging on the power of graph-cuts algorithms in clustering graphs. In our formulation the nodes are represented by people in the scene and the candidate o-space centres, while edges are defined between each pair of nodes of different type (i.e. between a person and a candidate o-space centre). We model the probability of each individual to belong to a specific o-space as a gaussian distribution. We then build the cost function by adding a Minimum Description Length (MDL) prior and considering the log function of the probability obtained. Moreover, we introduce an additive term which acts as the visibility constraint on the individual i regardless of the group person j is assigned to.

You can find more details about the method on the original PloS One paper or on the dedicated webpage.

GCFF is released under the MIT License (refer to the LICENSE file for details).

If you find GCFF useful in your research, please consider citing:

@article{setti2015f,
    title={F-formation detection: Individuating free-standing conversational groups in images},
    author={Setti, Francesco and Russell, Chris and Bassetti, Chiara and Cristani, Marco},
    journal={PloS One},
    volume={10},
    number={5},
    pages={e0123783},
    year={2015},
    publisher={Public Library of Science}
}

• MATLAB with a supported compiler

1. Clone the GCFF repository:

   $ git clone https://github.com/franzsetti/GCFF.git

   Lets call the directory you cloned GCFF_ROOT.

2. The repository already contains mex files for Windows and Linux 64bit machines. If you have a different system, you will need to mex the files again. To do that: open Matlab, move to the graphopt subfolder and run compile.m script.

   > cd GCFF_ROOT/graphopt
   > compile.m

3. Add the folder and subfolders to matlab path

   > addpath(genpath(GCFF_ROOT)) ;

In order to run the example provided and to reproduce the experiments of the original paper, you may want to download the datasets here.

Each dataset folder contains 3 files:

• 'features.mat' -> all the detections are stored in the variable 'features'; features{t} is an Nx4 matrix that stores all the detections at frame 't', columns are [ ID, x, y, orientation ]
• 'groundtruth.mat' -> groups ground truths, timestamp variable is needed to match with the features.mat file.
• 'settings.mat' -> parameter settings for the experiments (see the code for more details). This file is just an example and it could be replicated for different versions of experiments. Please use your own settings file if you want to run other experiments.

For any problem by using this code or with the datasets, please send an email to francesco.setti@univr.it .