The recognition of copies of paintings is not a simple task and during the last years many methods have been developed based on different techniques such as colors, shapes or contours identification. For this reason we spent the first week to figure out what are the most appropriate solutions according to the available time and to our capabilities. As a result some previously stated decisions have been changed because we realized that they would have complicated the project without notable impact on the results.
From the user to the machine
The following week was dedicated to find a way to represent digital pictures of paintings in a suitable and essential form for our future purposes. Since our main idea is to compare skeletons, we are firstly aimed at creating a system (in collaboration with another group) that allows the user to interact with the uploaded picture which will be processed later. To this end we propose an interface, implemented in Matlab (Fig.1), that helps the user to manually create skeletons by selecting the points corresponding to the vertices of a general rules-based structure (these rules are presented in the following subchapter).
At the beginning of the work we were planning to represent that structure using 15 points to identify all the important body parts in order to delineating an accurate shape of the object. It is clear that the increase of the number of points that are used causes new difficulties and implies the growing of computational complexity of the algorithm. However, in order to not overload the software with redundant calculations, the number of vertices has been limited to 6. This choice has been undertaken because of simplicity reasons and because the positions of some vertices (that correspond to the following groups: basin – hips, chest – shoulders – head, elbows – hands, feet – knees) can be grouped thanks to their closeness without losing efficiency of the comparing method. Therefore only one member of the mentioned groups will be considered.
Indeed the selected points for a canonical skeleton structure are the head, hands, basin and the feet. Regarding the general pattern identification (i.e. comparing the entire pictures themselves instead of specific details of them) the structure composed only by heads is still used.
The interaction between the user and the software will be performed by creating only single skeletons graphically while the general pattern of the loaded image will be automatically stored by clicking on the head button.
Storing input vertices
Each vertex will be categorized by a set of information (Cartesian coordinates, skeleton number, body part label) that will be used afterwards to perform the comparison. Since our main comparing concept is based on distances and angles, using a Cartesian reference frame to save selected points seems to be the most intuitive and effective way. Moreover, the use of an absolute reference frame will simplify the vertices overlapping procedure of our algorithm because the translations are univocally defined.
As soon as the interface is finished the following step will be the creation of the database of skeletons. This dataset will be the source of our comparison algorithms and it will allow us to guarantee the reliability of the codes. Defining and implementing an effective similarity recognition method is the crucial part of our project. Therefore the coding activity will be dealt for a considerable time once the database is composed.
Mattia Bergaglio, Alessio Santecchia, Daniyar Chumbalov