Frequency-based Motion Representation
for Video Generative Adversarial Networks

Supplementary Material
Sangeek Hyun, Jaihyun Lew, Jiwoo Chung, Euiyeon Kim and Jae-Pil Heo
[Paper]         [Code]

1. Toy Dataset

Removing scenario on toy dataset.

Original videos

Low frequency only videos

High frequency only videos

2. Randomly Sample Videos

Randomly generated videos on UCF-101, FaceForensics, VoxCeleb2-mini and BAIR-robot datasets.

3. Video Diversity

Each row indicates different motion codes with same content code.
Each row indicates same motion code with different content codes.

BAIR-robot

BAIR-robot

FaceForensics

FaceForensics


4. Temporal Interpolation

Interpolated videos on UCF-101, FaceForensics, VoxCeleb2-mini and BAIR-robot datasets.
Video interpolation on BAIR-robot dataset

Original videos (4x slow)

Interpolated Video
Video interpolation on FaceForensics dataset

Original videos (4x slow)

Interpolated Video
Video interpolation on UCF-101 dataset

Original videos (4x slow)

Interpolated Video
Video interpolation on VoxCeleb2-mini dataset

Original videos (4x slow)

Interpolated Video


5. Temporal Extrapolation

Extrapolated videos on UCF-101, FaceForensics, VoxCeleb2-mini and BAIR-robot datasets.
Extrapolated Videos (4x longer): the red box denotes extrapolated frames

UCF-101

Faceforensics

VoxCeleb2-mini

BAIR-robot


6. Frequency-based Motion Manipulation

Each column shares motion codes of low frequency components and each row shares motion codes of high frequency components.
Manipulated videos on BAIR-robot dataset

Manipulated videos on VoxCeleb2-mini dataset

Manipulated videos on FaceForensics dataset



7. Manipulation - Removing

Removing scenario on BAIR-robot dataset

Original videos

Low frequency only videos

High frequency only videos
Removing scenario on FaceForensics dataset

Original videos

Low frequency only videos

High frequency only videos
Removing scenario on VoxCeleb2-mini dataset

Original videos

Low frequency only videos

High frequency only videos


8. Manipulation - Altering

Each column indicates videos altered a specific frequency range of motions with the same content code.
The blue box denotes the original video.
Left column is altered low frequency components of motions.
Right column is altered high frequency components of motions.

BAIR-robot (low frequency altering)

BAIR-robot (high frequency altering)

FaceForensics (low frequency altering)

FaceForensics (high frequency altering)

VoxCeleb2-mini (low frequency altering)

VoxCeleb2-mini (high frequency altering)



Acknowledgements

This template was originally made by Phillip Isola and Richard Zhang for a colorful ECCV project; the code can be found here.