Code for feature extraction appears in Feature Extraction folder of Unity3D Assets.
For geometric features geometric_feature_extraction.cs provides a button in Unity editor and handles iterating over all scenes. For functional features the corresponding script is functional_feature_extraction.cs. write_scene_info.cs also provides a button 'Write Scene Info' in the editor --- this should be ran before any feature extraction.
Diagrams and descriptions of the features are provided in feature extraction.pdf.
Each feature is calculated between a pair of objects e1,e2. Features are not necessarily symmetric.
Functions for calculations are given in geometry_calculations.cs
For some geometric features meshes must be approximated using "Colliders" for computational reasons. This occurs in set_distance_properties(). If the mesh is can be appropriately approximated by it's convex hull, a box collider or a sphere collider this is used. Else, the mesh is approximated by a collection of box colliders using the NonConvexMesh Collider script.
Shortest distance between e1 and e2 is calculated by iterating over the vertices of e1 to find which vertex is closest to an approximation of e2. The shortest distance is then just the distance from this vertex to the approximation of e2.
containment (referred to as bbox_overlap_proportion) is calculated by first calculating the volume of overlap of the axis-aligned bounding boxes of e1 and e2. This volume is then divided by the volume of the axis-aligned bounding box of e1.
f_covers_g relies on a few variables.
Firstly height_seperation: if the bottom of e2 is above the top of e1 then the height seperation is height_separation = bottom of e2 - top of e1. Else, if the bottom of e1 is above the top of e2 then the height seperation is height_separation = bottom of e1 - top of e2. Else the height seperation is 0.
Next projection_angle: this is set to 5 degrees and accounts for the idea that as a covering object moves further from the object it is covering, it must be larger to provide the same covering function. Imagine an umbrella moving further and further away above somebody.
The projection angle and height seperation are then used to create the horizontal area that e1 is aiming to cover, new_e2_area --- the area of e2 plus the extra. The area of overlap of this area with the horizontal area of e1 is then calculated. This overlap area is then divided by the total horizontal area of new_e2_area.
g_covers_f is the same as f_covers_g but with e1 and e2 swapped.
horizontal_distance is just the horizontal distance between the centre of mass of e1 and the centre of mass of e2.
To calculate contact (referred to as contact_proportion), the number of vertices of e1 which are under some threshold distance to the approximation of e2 are counted. This is then divided by the total number of vertices in the mesh of e1.
To calculate above_proportion, the number of vertices of e1 which are above the highest point of e2 are counted. This is then divided by the total number of vertices in the mesh of e1.
To calculate below_proportion, the number of vertices of e1 which are below the highest point of e2 are counted. This is then divided by the total number of vertices in the mesh of e1.
As simulation is used to calculate these features, the functions for calculations are given in an in-script WaitingGame_FunctionalFeatureExtraction.cs
In order to the physics of falling/moving objects to work appropriately some objects need assigning an NonConvexMeshCollider approximation. All physics is handled in set_objects_at_rest.cs script.
Necessary assets are provided in Unity3D Annotation Environment folder, except NonConvexMesh Collider asset which must be purchased. This is used for better object physics.
These calculations are executed in print_support_positions().
Firstly the script gets the initial height of the centre of mass of the figure. Then the ground object is removed. Then the game waits 10 seconds and gets the height of the figure centre of mass. The raw movement is the initial position minus the new position.
This value is then normalized by an appropriate ground_height. If the bottom of the figure is above the top of the ground (within a threshold), then ground_height is just the height of the ground object. Else, if the bottom of the figure is above the bottom of the ground(within a threshold), then ground_height is fig_bottom - ground_bottom_point. Else, if the initial centre of mass is above the bottom of the ground(within a threshold), then ground_height is fig_com_y_initial - ground_bottom_point;. In all other cases ground_height is just the height of the ground object.
The normalized_movement is the raw movement divided by the ground_height. This is capped between 0 and 1.
These calculations are executed in print_lc_positions().
Four seperate calculations are performed for each cardinal direction. These values are then averaged in later processing.
In one cardinal direction a force is applied to the ground object. The horizontal movement in the direction of the force of the centre of mass of the figure is measured. This is then divided by the movement (in direction of force) of the centre of mass of the ground. This value is then capped between 0 and 1.