This is an initial commit implementing PatchBatch - a Batch Augmented Loss for Optical Flow by Dedi Gadot and Lior Wolf from Tel Aviv University (link), published at CVPR, 2016.
SPCI models are based on Optical Flow Requires Multiple Strategies (but only one network) by Tal Schuster, Lior Wolf and Dedi Gadot (link) published at CVPR, 2017.
PatchBatch achieved state-of-the-art results in 2016 on the KITTI (2012+2015) Optical Flow benchmarks and was ranked 6th on MPI-Sintel, though ranked 1st for small displacements.
The code was developed on Ubuntu 14.04, using Theano+Lasagne+OpenCV. You can see the performance it achieved on the KITTI2012, KITTI2015 and MPI-Sintel optical flow scoreboards.
- Clone this repository by typing:
git clone https://github.com/DediGadot/PatchBatch - Create a python (2.7) virtualenv, by typing:
virtualenv --no-site-packages env - Activate virtualenv: `source env/bin/activate'
- Install all the python packages described in Requirements.txt by typing:
pip install -r Requirements.txt - Make sure to configure Theano to your needs (GPU usage preferred - tested with OpenBlas and CuDNN 5.1)
- To check theano installation - run
python -c 'import theano; theano.test()'
The PatchBatch pipeline consists of the following steps:
- Input: two grayscale images, with the same shape
- Calculate descriptors (per each pixel in both images) using the PatchBatch CNN, i.e calculate a [h,w,#dim] tensor per image
- Find correspondences between both descriptor tensors using PatchMatch, with an L2 cost function
- Eliminate incorrect assignments using a bidirectional consistency check
- (Not yet implemented in this repository) Use the L2 costs + EpicFlow algorithm to interpolate the sparse optical flow field into a dense one (we used the default parameters of EpicFlow)
- Outputs: A->B optical flow field, (optional) descriptors file, cost assignment file
To run the PatchBatch pipeline, use the following syntax:
python patchbatch.py <img1_filename> <img2_filename> <model_name> <output_path> [optional -bidi] [optional --descs]
Currently supported models:
- KITTI2012_CENTSD_ACCURATE
- KITTI2015_CENTSD_ACCURATE
- KITTI2012_SPCI
- KITTI2015_SPCI
If the output_path does not exist, it will be created. In it will be placed the following:
- flow.pickle -
- A [h,w,3] numpy array with channel 0,1,2 being U, V, valid flag components of the flow field
- If the
-bidiflag is invoked, the code will compute 2 flow fields: img1->img2 and img2->img1 and will mark as 'invalid' all correspondences with inconsistent matchings (i.e. >1 pixels apart)
- cost.pickle -
- A [h,w] numpy array containing the matching cost per match
- (If the --descs option was used) descs.pickle -
- A list with two [h,w,#d] numpy arrays, the first contains descriptors per each pixel of img1, and the second the same for img2
You can also use benchmark_kitti.py to run a full benchmark on a folder with KITTI images.
For now, the EpicFlow extension is not yet implemented - so what you're getting is a pure PatchBatch descriptors + PatchMatch result.
The PatchBatch pipeline couldn't be achieved without the following great software pieces:
We also used the following toolkit for visualization:
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.