GitHub - LiZimo/FuncFlow: Unsupervised Cosegmentation and Optical Flow fields using functional maps

LiZimo / FuncFlow Public

Notifications You must be signed in to change notification settings
Fork 2
Star 4

Unsupervised Cosegmentation and Optical Flow fields using functional maps

Notifications

Name		Name	Last commit message	Last commit date
Latest commit History 21 Commits
data/iCoseg/icoseg/skate2		data/iCoseg/icoseg/skate2
external		external
src		src
demo.m		demo.m
demo_func.m		demo_func.m
example_output.jpg		example_output.jpg
funcflow_report.pdf		funcflow_report.pdf
license.txt		license.txt
readme.txt		readme.txt

Repository files navigation

==============================================
README
funcflow v 0.9

Copyright (c) 2015, Zimo Li
All rights reserved.

==============================================
Work in Progress
==============================================
This is an implementation for the funcflow framework.  For 
details on the method, please refer to the pdf attached in this folder.  
The code has been tested on Matlab 2015a on a CentOS 7 machine.

In short, the algorithm initializes correspondence among a network of images
using an off-the-shelf optical flow algorithm such as SIFTflow(http://people.csail.mit.edu/celiu/SIFTflow/) projected into a reduced
space.  Then, through alternating optimization, we optimize new correspondences 
that are consistent across the network of images.  We optimize in the reduced space
using functional maps, as it is easier to do so.  The final output are 
consistent segmentations across the image collection, as well as pairwise
optical flow maps derived from the final maps.

For details on the method,
refer to the attached PDF in the top level directory.

The code works in both the pixel and superpixel domains.  I have 
found that working in the pixel domain has better results.

=============================================
Dependencies:
=============================================
cvx (http://cvxr.com/cvx/)
dsp (http://vision.cs.utexas.edu/projects/dsp/)
gbvs (http://www.vision.caltech.edu/~harel/share/gbvs.php)
gist (http://people.csail.mit.edu/torralba/code/spatialenvelope/) 
gop (http://www.philkr.net/home/gop)
siftflow ( http://people.csail.mit.edu/celiu/SIFTflow/ )
SLIC (http://ivrl.epfl.ch/research/superpixels) 

Please download/install all these packages in the "external" folder
The ones with public licenses have already been added (cvx, gbvs, gop, SLIC).  

For additional datasets, please download iCoseg(http://chenlab.ece.cornell.edu/projects/touch-coseg/)
and MSRC (http://research.microsoft.com/en-us/projects/objectclassrecognition/)

==============================================
Running the code:
==============================================
To run the pipeline, simply execute demo.  The code assumes the images you
want segment are all in the same directory.  For evaluation, it assumes there is a 
sub-directory in your image-directory named "GroundTruth" that has the ground truth
binary masks of every photo, with the same name.  

There are a number of parameters with explanations in the demo.m file 
that can be modified.


==============================================
Additional Tools:
==============================================

There are a number of scripts in "funcflow/src/visualizations" that are not
used in the pipeline, but which produce figures of intermediary steps
that might be of interest (such as some of the ones produced in the pdf).