Effect of DTW temporal window - 1

 

The following example shows the effect of the DTW window size on the match results. The matching window size is increased from 1 to 40 keyframes. That means the dynamic warping algorithm will look for the closest point that is temporally +δ or -δ keyframes (in the match) away from the current keyframe (in the query), where δ is the window size. If δ=1, this is equivalent to Euclidean matching. We show the returned matches for each window size.

The query is a short sitting motion 53 keyframes long (or 2.2 seconds).

Matching parameters: DTW, rotational parameterization, full-body area, no global position, no global rotation.

 

DTW Example 1

 

Video Description

(in order of appearance)


Full-size video

  1. The query
  2. Long pause
  3. δ=1. The  best 5-matches, from best to worst.
  4. Long pause
  5. δ=5. The  best 5-matches, from best to worst.
  6. Long pause
  7. δ=10. The  best 5-matches, from best to worst.
  8. Long pause
  9. δ=20. The  best 5-matches, from best to worst.
  10. Long pause
  11. δ=40. The  best 5-matches, from best to worst.

 

 

 

Effect of DTW temporal window - 2

 

Same as above except that the query motion is a slow walking motion 56 keyframes long (or 2.3 seconds).

Matching parameters: DTW, rotational parameterization, full-body area, no global position, no global rotation.

 

DTW Example 2

 

Video Description

(in order of appearance)


Full-size video

  1. The query
  2. Long pause
  3. δ=1. The  best 5-matches, from best to worst.
  4. Long pause
  5. δ=5. The  best 5-matches, from best to worst.
  6. Long pause
  7. δ=10. The  best 5-matches, from best to worst.
  8. Long pause
  9. δ=20. The  best 5-matches, from best to worst.
  10. Long pause
  11. δ=40. The  best 5-matches, from best to worst.