This code implements the entire Attentive Module of
"On the use of efficient projection kernels for motion-based visual saliency estimation." Nicora, Elena, and Nicoletta Noceti. Frontiers in Computer Science: 67.
"Efficient Projections for Salient Motion Detection and Representation" Nicora Elena, Università degli Studi di Genova (2022)
The algorithm requires in input a pair of maps, max_pooling and avg_pooling, and returns a refined segmentation of the moving object in the scene.
IMPORTANT: Pooling maps required in input are computed using 3D Gray-Code Kernels (see paper cited above for more information)
For each pair of pooling maps we can identify the following steps:
- Otsu's adaptive thresholding: returns a coarse segmentation (GCKsA) based only on max_pooling
- Refinement of GCKsA using minima and maxima of the avg_pooling: here we discard connected components of GCKsA smaller than a 5x5 patch and those that do not correspond to areas where there are local extrema
- Identification of "present" and "past" values by using two thresholds (delta_1 and delta_2) over the avg_pooling
- Composition of the refined segmentation (GCKsR) subtracting the pixels corresponding to the "past" and adding the "present" pixels to consolidate the final result.
Note that every step contains several morphological operations in order to produce a visually pleasing segmentation with the least number of holes and artifacts.
Users can tune variables delta_1 and delta_2.
We assume that values or the avg_pooling near zero correspond to the past positions and values near 1 correspond to the present. But notice that in some cases it is the opposite so the user needs to invert the name of the variables in first_ref (first refinement that subtracts the "past") and second_ref (second refinement that adds the "present").