This MATLAB software accompanies our TPAMI'14 and ECCV'12 papers:
K. Karsch, C. Liu and S.B. Kang. "DepthTransfer: Depth Extraction from Video Using Non-parametric Sampling." IEEE TPAMI, 2014.
K. Karsch, C. Liu and S.B. Kang. "Depth Extraction from Video Using Non-parametric Sampling." ECCV 2012.
Descriptions of parameters and implementation details can be found in the paper and supplementary material (available at https://kevinkarsch.com). Please cite the TPAMI paper if you use our code for your research.
Microsoft inadvertently removed several of their datasets and research pages during a recent server change, including MSR-V3D, and I have been told that it is unlikely ever be replaced. However, there are other datasets which are all-in-all better, serve similar purposes:
Pre-mexed binaries for Windows (32b and 64b), OS X (64b), and Linux (32b and 64b) are packaged with the code. If you're using any of those OSes, you should not need to compile anything. These binaries and code have been verified with MATLAB 7.14 (2012a).
If you need to compile the code (either on a different OS or the binary isn't suitable for your version of MATLAB), all you need to do is run 'make' from a MATLAB prompt. If you run into any errors, see make.m.
IMPORTANT: to use optical flow (motion estimation + temporal consistency),
you must also download Ce Liu's publicly available optical flow code. Modify the path in opticalflow.m (OPTICAL_FLOW_DIR) to match the location
that you unpacked the OpticalFlow directory (the default is to unpack to
this directory). Note that if you are only running on single images or on
data from our dataset (StereoRGBD Dataset), this is not necessary since the
flow has been precomputed.
To get started, take a look at demoDepthTransfer.m, or just run 'demoDepthTransfer' from a MATLAB prompt. This runs depth transfer on a single image from the Make3D Range Image Dataset.
To evaluate our code on the Make3D dataset, first read the documentation in evaluateMake3dDataset.m and run 'evaluateMake3dDataset' from a MATLAB prompt.
For more advanced usage, there are several scripts in the examples directory (these require the optical flow module; see Installation above):
example1.m - Demonstrates depth transfer on a static video sequence with
moving objects
example2.m - Demonstrates depth transfer on a rotating video sequence
example3.m - Demonstrates depth transfer on a sequence with varying
focal length and moving objects
example4.m - If you have also downloaded our StereoRGBD dataset, this
example will run one of the results using our training/test
split.
DepthTransfer/ - Root directory
README.txt - this file
make.m - Helper function to compile SIFTflow and other
cpp-mex files
createData.m - Stores raw data in DepthTransfer format
loadData.m - Reads DepthTransfer data stored on disk
initializeProject.m - Sets up a DepthTransfer project (params, paths,
etc)
depthTranser.m - main DepthTransfer module (relies on code in
matlab/)
demoDepthTransfer.m - Simple demonstration of the code on a single image
evaluateMake3dDataset.m - reproduces Make3D results from our paper
opticalflow.m - A wrapper for Liu's optical flow (see above)
fillDepthHoles.m - Code to spatiotemporally interpolate holes in
depth maps
demo/ - Scripts and data for running advanced
DepthTransfer demos
matlab/ - Directory containing core DepthTransfer code
Features/ - Code for computing image and flow descriptors
Inference/ - Code for inferring depth from warped
candidate depth images
MotionEstimation/ - Several algorithms for estimating a binary
motion segmentation in videos
SIFTflow/ - A modified version of Liu et al.'s SIFTflow**
Utilties/ - DepthTransfer helper functions
examples/ - Example scripts and data
data/ - Default data directory (empty initially)
results/ - Default results directory (empty initially)
**Original SIFTflow code and paper can be found here:
http://people.csail.mit.edu/celiu/SIFTflow/
This code is not optimized for efficiency in any sense, and can take both a very long time to run and a great deal of memory. Take advantage of MATLAB's parallel toolbox if available (ie. 'matlabpool'), and be careful about the size of the input sequence. Using a project size of height=300 and width=200, a video of 60 frames can take several GB of RAM.
For bug reports, please send me an email at [email protected]