Emotion Control of Unstructured Dance Movements

 

Supplementary Materials:

It is difficult to fully represent by-products of our results, and appreciate the quality of modified dance motions in the paper. Therefore, in these pages we include larger format versions of the intermediate results in the paper, together with more dance editing results, as well as our user study analysis. Furthermore, we present the results of the Yumer and Mitra [2016] method. We recommend using the latest version of Firefox, Chrome, and IE to view these results.

1.    Clustering Results:

We use the hierarchical clustering algorithm to group similar poses into clusters, resulting in five clusters (crouching, jumping and three different standing styles). Here we show 100 randomly selected exemplar poses for each cluster.

Cluster 1 (standing poses) , Cluster 2 (crouching poses) , Cluster 3 (standing poses) , Cluster 4 (jumping poses) , Cluster 5 (standing poses)

 

  1. Local emotion-content tables on all features

 

  1. Local emotion-content tables on EFFECTIVE/CONSISTENT features

 

  1. Additional dance stylization results:
    Note: while viewing our results, please recall that our method modifies the input motion toward a target emotion; it does not replace the emotion. Therefore, an input motion becomes “more happy”, or “more sad”, etc.

 

Input Motion

Emotion Modification

Results

Dance-1

more happy

+Happy

Dance-1

more sad

+Sad

Dance-1

more tired

+Tired

Dance-2

more excited

+Excited

Dance-2

more miserable

+Miserable

Dance-3

more excited

+Excited

Dance-3

more sad

+Sad

Dance-4

more happy

+Happy

 

 

5.    User Study:

Representative screen shots for each section of our user study.

       Survey 1

General Questions

Section 1: Emotion Recognition

Section 2: Emotion Modification

Section 3: Motion Realism

 

 

       Survey 2

Motion Realism

 

 

6.    Comparison with Yumer and Mitra [2016]:

The most closely related prior work that transfers styles on independent actions and can be used for comparison is Yumer and Mitra [2016]. The authors kindly helped in the evaluation procedure, running our heterogeneous dance motion data in their system. However, since our dance motions are highly non-regular and arbitrary, and since their system is trained using motions of different nature (locomotion), the produced stylized motions suffer from abnormal movements and foot-skating.

Yumer and Mitra [2016] method is limited to transfer style only to those emotions that are in the training dataset, thus style was transferred only to Angry and Depressed (Sad) emotions.

·        Input video

·        Angry video

·        Depressed video

 

 

References

 

YUMER, M. E., AND MITRA, N. J. 2016. Spectral style transfer for human motion between independent actions. ACM Trans. Graph. 35 (Aug.).