This software package includes tools for filtering noisy clicks in an application of interactive object segmentation. The intented clicks are collected from an online and crowdsourcing campaign.

The software [Download ZIP (151M)] was developed by Ferran Cabezas during his BSc thesis. Read it for more details. 

Folders:

• data:

• results_1st_round: Load the results from the Bounding Box experiment: we had several users draw a bounding box around the object

• results_2nd_round: Load the results from the crowd  experiment on Click'n'Cut

• results_1st_round_cleanUp: users from results_2nd_round that have FULLY realised the experiment

• results_experts: Load the results from the Expert experiment on Click'n'Cut

• task:  This structure contains information about the 105 tasks that we study id_task ; image name ; description of the object ; mask name. Tasks 1-100 are considered as a test set, and Tasks 101-105 are considered as a train set.

• Felz_Superpixels: Felzenzwalb superpixel oversegmentations with sigma=0.5, m=20 and different values of 'k'. In the file 'readme' it is  shown the correspondences of the folders and the 'k' values.

• GTmasks: Ground truth masks of all images that are in the folder 'Images'

• Images: Set of images from PascalVoc and DCU-Berkeley dataset.

• MCG: It is conatined the MCG(candidates) masks from all tasks

• Segmentation:

• computeJaccard: It is computed the Jaccard index given two different masks.

• filterMasks: We use the median bounding box from the Bounding Box experiment to filter out some of the masks.

• findBestMask: It is given to each mask a score and it is kept the masks that have higher score.

• Superpixels_Achante: Achanta(also called as SLIC) superpixel oversegmentations with regularizer=0.1 and different values of 'RegionSize'. In the file 'readme' it is  shown the correspondences of folders and 'RegionSize' values.

• Superpixels: N-cuts superpixel oversegmentation.

Functions:

• filterClicks: Given a set of users it is returned two sets using both partial and total filtering clicks techniques.

• MCG_RemoveUsers: Jaccard index calculation in the test set given  different set of users sorted by its error and jaccard indexin the train set. The mask used to compute the Jaccard index in the test set it is obtained by using precomputed object candidates(this is why in the function appears 'MCG' as it is reffered to the object candiates).

• MCG_removeUsers_FilterClicks: Same functionality as 'MCG_RemoveUsers' but each time it is taken a different set of users, it is filtered the clicks with both techniques from filtering them: partial and total.

• probMap_errorRate: Foreground map creation given the error rate in the train set.

• probMap_JaccIndex: Foreground map creation given the Jaccard index in the train set.

• users_categorization: Given the manual rules, it is calculated from the test set all 6 features for each user in order to automatic categorize them.

Variables: 

• UserErrorsGS: Error rate from each user in the train set

• UserJaccardGS: Jaccard index from each user in the train set.