Human Motion: How to Evaluate Segmentation w.r.t. Retrieva
Michal Balazia
Laboratory of Data Intensive Systems and Applications Faculty of Informatics Masaryk University
May 5, 2014
Michal Balazia  (DISA) Human Motion Segmentation May 5, 2014 1/8
The Goa
We have the following segmentation technique:
Time
Our goal: Use this segmentation to improve retrieval How?
Create a motion index at the level of segments What can be the problems?
• representation of segments
• distance function
• clustering
Michal Balazia (DISA)
Human Motion Segmentation
May 5, 2014
2/8
Evaluation Methods
1) Segment-Shift Retrieval
2) Quality of Clustering
3) Segment Positions
4) Distance (Order) Preserving
Segment-Level Motion Retrieva
Ground truth provided - by exhaustive pose-level search What can segments represent?
• search shift units (Segment-Shift Retrieval)
a process all segments - not very efficient
• retrieve all occurences of Q in DB - most effective
• words (Text-Based Retrieval)
• process only some segments - much more efficient
• retrieve almost all occurences of Q in DB - still enough effective
A typical text-based retrieval:
1) Partition input motion data into segments
2) Cluster those into groups to treat similar segments as equal
3) Mark the clusters with letters
4) Use a text-based searching technique to retrieve occurences of Q in DB
Michal Balazia  (DISA) Human Motion Segmentation May 5, 2014        4 /8
Evaluation Method 1: Segment-Shift Retrieva
DB:
Q:
length variation (l|,2n)
cLgt ■■■ distance to the fc-th closest occurence of Q in ground truth
dg ... distance to the fc-th closest occurence of Q with use of segmentation
1 -00.0
i/lotion Segmentation
Evaluation Method 2: Quality of Clustering
DB: AABEDAEC... Q: DAE
Michal Baläzia (DISA)
Human Motion Segmentation
May 5, 2014
6/8
Evaluation Method 3: Segment Positions
ground truth
DB: h
i i i i        i i i i i i i h
bcd bcdaad        c bc
Q. i i i i i •   a b c d
ground truth
DB:
ab ab        ab ab
Q:
i i i
a b
i/lotion Segmentation
Evaluation Method 4: Distance (Order) Preserving
motion M
compressed motion c(M) - only key poses (determined by segmentation) Distance preserving:
d(MuM2) = d(c(M1),c(M2)) Distance order preserving:
d{M1,M2) < d{M3,M4)     d{c{M1),c{M2)) < d{c{M3), c(M4))
me m4 m7 m,
m:
m-
c(m5; c(m7; (m4; c(m5; c(m3; c(m2;
1
i/lotion Segmentation