Awesome

Code under development. Keep updated to the latest version.

The 2017 DAVIS Challenge on Video Object Segmentation

Package containing helper functions for loading and evaluating DAVIS.

A Matlab version of the same package is also available.

Terms of Use

DAVIS is released under the BSD License (see LICENSE for details)

Introduction

DAVIS (Densely Annotated VIdeo Segmentation), consists of high quality, Full HD video sequences, spanning multiple occurrences of common video object segmentation challenges such as occlusions, motion-blur and appearance changes. Each video is accompanied by densely annotated, pixel-accurate and per-frame ground truth segmentation.

Code Usage

Evaluate

In order to evaluate your results, execute the script ROOT/python/tools/eval.py providing the segmentation and setting the correct phase (train,val etc...) and year (2016,2017).

• The evaluation on the 2016 dataset can be done either as single object (ver. 2016) or on multiple objects (ver. 2017, default). For single object evaluation set the flag: -single-object.

Evaluate your results on the {year} dataset (ver. 2017, multiple objects):

python tools/eval.py -i path-to-my-technique -o results.yaml --year {year} --phase val

Evaluate your results as in 2016 with single objects (ver. 2016)

python tools/eval.py -i path-to-my-technique -o results.yaml --year 2016 --single-object --phase val

Read and write

Annotations are provided in PNG indexed format. We provide helper functions to read and write individual images:

• davis.io.imread_indexed(filename)
• davis.io.imwrite_indexed(filename,array)

Note that for a correct evaluation of the results, it is critical to preserve the same labeling of the annotations, i.e. assign identical numbers to corresponding objects.

See experiments/read_write_segmentation.py for usage examples

Visualize

Execute the script ROOT/python/tools/visualize.py. The command-line arguments are similar to the evaluation script. Use --single-object to visualize the original DAVIS 2016.

Example: python tools/visualize.py -i path-to-my-technique --year 2017 --phase val

Dependencies

C++

• Boost.Python

Python

• See ROOT/python/requirements.txt (Optionally to visualize results install cv2)

Installation

C++

1. ./configure.sh && make -C build/release

Python:

1. pip install virtualenv virtualenvwrapper
2. source /usr/local/bin/virtualenvwrapper.sh
3. mkvirtualenv davis
4. pip install -r python/requirements.txt
5. export PYTHONPATH=$(pwd)/python/lib
6. See ROOT/python/lib/davis/config.py for a list of available options

Documentation

The directory is structured as follows:

• ROOT/cpp: Implementation and python wrapper of the temporal stability measure.

• ROOT/python/tools: contains scripts for evaluating segmentation.

  • eval.py : evaluate a technique and store results in HDF5 file
  • eval_view.py: read and display evaluation from HDF5.
  • visualize.py: visualize segmentation results.

• ROOT/python/lib/davis : library package contains helper functions for parsing and evaluating DAVIS

• ROOT/data :

  • get_davis.sh: download input images and annotations.

See source code for API documentation.

Download Dataset

We provide a script to convenitently download the data. Simply execute: sh ROOT/data/get_data.sh. Alternatively the dataset can be downloaded from the DAVIS Challenge website.

Citation

Please cite DAVIS in your publications if it helps your research:

@inproceedings{Perazzi_CVPR_2016,
  author    = {Federico Perazzi and
               Jordi Pont-Tuset and
               Brian McWilliams and
               Luc Van Gool and
               Markus Gross and
               Alexander Sorkine-Hornung},
  title     = {A Benchmark Dataset and Evaluation Methodology for Video Object Segmentation},
  booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2016}
}

@article{Pont-Tuset_arXiv_2017,
  author  = {Jordi Pont-Tuset and
             Federico Perazzi and
             Sergi Caelles and
             Pablo Arbel\'aez and
             Alexander Sorkine-Hornung and
             Luc {Van Gool}},
  title   = {The 2017 DAVIS Challenge on Video Object Segmentation},
  journal = {arXiv:1704.00675},
  year    = {2017}
}

Contacts

• Federico Perazzi
• Jordi Pont-Tuset

TODOs

• Temporal stability measure (T)
• Per-attribute evaluation script
• Add usage examples