estimating_variation

Discussions of commit f0d590c3d384f8ca92dd2b4ece7a15f54bc78793

Amelie and Simon

• Variation of bounding boxes
  • Functional: (i) there will be high variation geometrically; (ii) actually there will be low variation geometrically because objects are much more restricted to particula robjects and therefore restricted to particular locations
  • Bounding boxes:
    • Normalisation against the image dimensions; the same two objects would be spatially very different if the image is taken from close and from far
    • Currently, normalisation ensures that all images have the same dimension
    • We project the x, y. w, h into a mask, i.e. a vector of 0 and 1 (49 locations)
    • To do:
      • To plot similar graphs for every target-landmark pair; what do they look like?
  • Estimate the similarity between two graphs using cosine?
  • Estimate the variation of targets/landmarks of a particular relation and then rank all prepositions by this variation
  • What variation?
    • Currently stdev is calculated for the entire x, y, h, w: hence also on the height and widths of objects; but these are the same between the relations
    • Solution 1: create a mask with targets (or landmarks); compare targets pairwise with cosine; take the stdev for cosine

     tar1  tar2 tar3
tar1  *    *    *
tar2       *    *
tar3            *

* Visual similarity
  - For each relation extract visual features of targets and visual features of landmarks prepositions_bboxes[p].append([v_target, v_landmark])
  - For every preposition p calculate cosine similarity between the visual features of every v_target (and the same for for v_landmark)
  - Estimate the variation of cosine for v_tragets
  - Is there a difference between relations, i.e. are targets fo fucntional relations more similar than those of geometric ones?