Clustering of similar motions in the given “Bharatanatyam” dance video.
-
The video was given in RGB format
-
Annotation file with motion frame details.
- Cluster number for each motion
Adavu | Performance1
total frames |
Performance 2
total frames |
Performance 3
total frames |
Total motions | no of unique motions |
Natta1 | 1546 | 1590 | 1532 | 32 | 4 |
Natta2 | 1557 | 1522 | 1545 | 32 | 4 |
Natta3 | 2680 | 2698 | 2760 | 64 | 8 |
Natta4 | 5537 | 5531 | 5504 | 128 | 8 |
Natta5 | 2580 | 2728 | 2748 | 64 | 10 |
Natta6 | 2781 | 2764 | 2729 | 64 | 12 |
Natta7 | 2828 | 3022 | 2706 | 64 | 14 |
Natta8 | 2710 | 2811 | 2752 | 48 | 11 |
-
All the motions didn’t have same no of frames even the similar ones.
-
Also the frames of similar motions may be different.
-
To compare similarity of two motions,need to find a good measure.
-
The similarity measure should be more for similar motions and less for different motions
-
It should be able to compare motions with different no of frames
-
Used Dense optical flow and obtained the feature vector of each motion.
- Tried various variations using HOF
-
Applied DTW as similarity measure for motions and obtained similarity matrix.
-
Used Spectral Clustering to cluster the data using the above similarity matrix.
-
Used Dense optical flow to get the features for each motion.
-
Made all vectors equal size by appending with small value(1e-5)
-
Used SVM for the classification