SpeedNet: Learning the Speediness in Videos

Supplementary Material

 

Additional Video Retrieval Results

We query the UCF101 training set by comparing SpeedNet embeddings of example clips from the UCF101 test set to all SpeedNet embeddings of the clips in the training set. Shown here is a diverse set of query actions, and their top-3 nearest neighbors.
Notice that although most of the retrieved videos are not of the same class, many of them contain motion patterns that are similar to the query. Since SpeedNet was trained solely to distinguish between natural (normal-speed) motion and sped-up motion, attenuation to type/speed/pattern of motion is plausible behavior.

 

  Query clip   Retrieved top-3 results
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   

 

Back to main supplementary material page.