For the reproducibility challenge we tried to reproduce the paper from Ranjan et al. Learning to count everything. This is the repository, where you can find our final report (not yet available), authors code, and also our own scripts.

The structure of the repository is as follows: in folder src there is authors original code, together with two other folders -- notebooks and data. Notebooks folder includes 11 notebooks, which are self explanatory. In the data folder, you have to put density maps provided by the authors in gt_density_map_adaptive_384_VarV2. Density maps that are provided separately (not together with the images) must be put in gt_density_map_special. Images must be put in images_384_VarV2. These are the setting for FSC-147 data set provided by the authors of the original paper and can be accessed here: https://github.com/cvlab-stonybrook/LearningToCountEverything.

Please put instances_val_2014.json and instances_val_2017.json which can be found on COCO data set website into src/notebooks/misc directory (only required for running COCO related notebook).

For this challenge we were running test_extended.py and notebooks. We also changed utils.py.

The majority of our experiments can be ran using script test_extended.py, which is only an extension of script test.py, produced by the authors of original paper. The instructions for running their script are described on their repository. We added some command line parameters in order to evaluate FamNet in different test scenarios. The added parameteres are:

• o: the path to the output file, in which the results are written,
• al: if specified, the absolute loss is used in MinCountLoss instead of squared loss,
• ne: the number of exemplars that is used (for number of exemplars ablation study), set to -1 by default, meaning that all exemplars are used,
• c: if specified, FamNet is evaluated on CarPK test set,
• cc: if specified, FamNet is evaluated on JHU-CROWD++ test set.

To evaluate different values for $ \sigma_G $ in perturbation loss, we ran script choose_sigma.py and to find the instances, where adaptation brings biggest/worst improvements we ran script check_adapt.py. Both of these scripts are ran in the same way as original test.py and have the same command line parameters (except for parameter -a, which is removed in both scripts and -sp, which is removed in choose_sigma.py).