Repository for different network models related to flow/disparity (ECCV 18)
Clone or download
Latest commit 204add3 Nov 13, 2018
Type Name Latest commit message Commit time
Failed to load latest commit information.
DispNet3 Remove duplicate net instances Aug 8, 2018
FlowNet3 Remove processid files Aug 8, 2018
FlowNetH updated download script for Pred-Merged Oct 9, 2018
SceneFlow Update Nov 12, 2018
.gitignore Add SceneFlow models Nov 12, 2018
LICENSE.txt Add license Aug 8, 2018 updated readme Sep 24, 2018


Repository for different network models related to flow/disparity from the following papers:

NOTE: We only provide deployment code for these networks. We do not publish any training code and also do not offer support about questions for training networks.

  • Occlusions, Motion and Depth Boundaries with a Generic Network for Disparity, Optical Flow or Scene Flow
    (E. Ilg and T. Saikia and M. Keuper and T. Brox published at ECCV 2018) [paper] [video]

  • Uncertainty Estimates and Multi-Hypotheses Networks for Optical Flow
    (E. Ilg and Ö. Cicek and S. Galesso and A. Klein and O. Makansi and F. Hutter and T. Brox published at ECCV 2018) [paper] [video]


Running networks

  • Change your directory to the network directory (Eg: FlowNet3)
  • Run This takes a while to download all snapshots
  • Now you should be ready to run the networks. Change your directory to a network type (Eg: css). Use the following command to test the network on an image pair: python3 eval image0_path image1_path out_dir

Output formats

The networks are executed using the scripts in the respective folders. Just running this controller will produce several output files in a folder (note that you can also obtain this output just as numpy arrays and write it to some custom files; see next section).

For optical flow we use the standard .flo format. The other modalities use a custom binary format called .float3. To read .float3 files to numpy arrays, please use the module.

Example usage:

from import read 
occ_file = 'occ.float3'
occ_data = read(occ_file) # returns a numpy array

# to visualize
import matplotlib.pyplot as plt

Controller eval

The eval method of the controller writes to the disk by default. To avoid writing to disk, create a Controller object and use the eval method available in the net_actions member variable. This can be useful if you want to process the output of our networks in memory and not incur additional disk I/O.

Example usage:

import netdef_slim as nd
c = Controller() 
out = c.net_actions.eval(img0, img1)
# out is an OrderedDict with the following structure
#OrderedDict(['flow[0].fwd',     np.array[...],
              'occ[0].fwd',      np.array[...],
              'occ_soft[0].fwd', np.array[...],
              'mb[0].fwd',       np.array[...],
              'mb_soft[0].fwd',  np.array[...],


netdef_models is under the GNU General Public License v3.0