Skip to content
master
Switch branches/tags
Go to file
Code

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
img
 
 
 
 
src
 
 
 
 

README.md

On-the-fly optical flow training data generation

This repository contains a data-generation layer for Caffe (v1). The layer takes a texture database, and produces optical flow training data (images and groundtruth) live at training time. It was used in the IJCV 2018 paper "What Makes Good Synthetic Training Data for Learning Disparity and Optical Flow Estimation?" (http://dx.doi.org/10.1007/s11263-018-1082-6), where it proved to produce better data (benchmarked on Sintel) than the original FlyingChairs dataset. You could (if you were fond of terrible puns) call this one On-the-Fly-ingChairs.

Predefined options include various object shapes and motion types.

Texture database must be provided by user.

Author

Nikolaus Mayer (mayern@cs.uni-freiburg.de)

License

This code is provided for research purposes only and without any warranty. Any commercial use is prohibited. If you use the code or parts of it in your research, you should cite the aforementioned paper:

@Article{MIFDB18,
  author       = "N. Mayer and E. Ilg and P. Fischer and C. Hazirbas and D. Cremers and A. Dosovitskiy and T. Brox",
  title        = "What Makes Good Synthetic Training Data for Learning Disparity and Optical Flow Estimation?",
  journal      = "International Journal of Computer Vision",
  number       = "9",
  volume       = "126",
  pages        = "942--960",
  month        = "Sep",
  year         = "2018",
  note         = "https://arxiv.org/abs/1801.06397",
  url          = "http://lmb.informatik.uni-freiburg.de/Publications/2018/MIFDB18"
}

The codebase utilizes the wonderful libraries CImg (download it and put CImg.h into include/thirdparty) and Anti-Grain Geometry.

Usage

See example/train.prototxt. The layer includes a number of hardcoded data modes, but is trivial to extend. Textures are provided as a list of image files (loaded once at startup).

The data dimensions are hardcoded in include/caffe/data_generation/DataGenerator.h, lines 55 et seq.

Note that this layer uses a lot of CPU power.

About

Caffe(v1)-compatible codebase to generate optical flow training data on-the-fly; used for the IJCV 2018 paper "What Makes Good Synthetic Training Data for Learning Disparity and Optical Flow Estimation?" (http://dx.doi.org/10.1007/s11263-018-1082-6)

Topics

Resources

Releases

No releases published

Packages

No packages published