multi-sepconv-interp

Video frame interpolation using multiple separable convolution

Data preprocessing

pyflow: Fast optical flow estimation.
My fork of pyflow: No messy logs / warnings that are printed to stdout.

• Youtube videos
• UCF Datasets:
  • UCF Sports Action Data Set
  • UCF YouTube Action Data Set
• KITTI Scene Flow 2015

UCF dataset's quality is too low and contains watermark so not suitable. KITTI's images are stretched / warpped so not suitable.

So my data are only sampled youtube videos.

Scripts

youtube-dl

Useful tool to batch download youtube videos.

https://github.com/rg3/youtube-dl

sample-video.py

Sample patch triples from videos.

list-valid-triple.py

Scan a directory and print a list of triples.

color-histogram.py

Accept a list of triples (output of list-valid-triple.py) and a threshold value. Filter out triples that potentially contains video shot boundary.

ref: http://www-nlpir.nist.gov/projects/tvpubs/tvpapers03/ramonlull.paper.pdf

texture-entropy.py

Filter out triples that has low texture entropy.

ref: http://scikit-image.org/docs/dev/auto_examples/filters/plot_entropy.html

flow-estimate.py

Estimate flow magnitude between first and third frame for each data.

flow-guided-sample.py

Flow-guided heuristically sample data.

plot-distribution.py

Plot histogram of distribution of data from the result of flow-estimate.py and flow-guided-sample.py.