Implementation of the Deep-Occlusion Reasoning

Paper: https://arxiv.org/abs/1704.05775 Project page: https://pierrebaque.github.io/page-DeepOcclusion/

Dependencies

The main dependency is the Theano package.
You also need to install the theano ROI-Pooling Layer from https://github.com/ddtm/theano-roi-pooling to get the unary potentials to work.

How to run.

Nb: The demo provided here is temporary and should be improved soon to make it more robust and easier to run. Happy to read your feedbacks.

Clone the github code

Download this zip file https://drive.google.com/file/d/0BxijKKbgC7wRTE03ajEtZ1FYc2M/view?usp=sharing

Unzip it in the root directory

From the same root directory, type:

mv tozip/models VGG/models
mkdir Potentials/Parts/Run
mkdir Potentials/Unaries/Run

Then, run the notebooks in the following order:
RunParts.ipynb
RunUnaries.ipynb
RunPom.ipynb

You should get some results. This is for running the pretrained model and not for training yet.

Running on your own dataset

In order to run the code on your own dataset, you just need to generate the ".pom" file corresponding to your cameras. This file describes the 2D boundary of each 3D bounding-box, projected on each camera, approximated by a rectangle. In order to generate such file, you can use our code with your camera calibrations: https://github.com/pierrebaque/generatePOMfile