Disclaimer: This is not an official release. This implementation is based on the ICRA 2017 paper of the same title by Sen Wang, Ronald Clark, Hongkai Wen, and Niki Trigoni. We try to reproduce the results presented in the above paper, while incorporating our own interpretations of the approach, wherever needed.

Implementation by: Krishna Murthy and Sarthak Sharma

Pretrained models and results will be pushed in due course of time.

This is a PyTorch implementation. We assume PyTorch and dependencies are setup. This code has been tested on PyTorch 0.4 with CUDA 9.0 and CUDNN <VERSION>.

Dependencies: scipy, scikit-image, matplotlib, tqdm, and natsort. We also use tensorboardX for visualization purposes.

scipy, scikit-image, matplotlib, and natsort can be installed using standard pip or conda commands.

pip install --user tqdm

conda install -c conda-forge tqdm

pip install tensorboardX

pip install git+https://github.com/lanpa/tensorboard-pytorch

To run the code, from the base directory of the repository, run something like this:

python -B main.py -datadir ~/scratch/KITTIOdometry/dataset/ -cachedir ../../scratch/DeepVOCache -nepochs 2 -tensorboardX True -lrScheduler cosine -expID tmp -scf 2 -lr 1e-3 -beta1 0.7 -momentum 0.009 -optMethod sgd -dropout 0.5 -modelType flownet -loadModel cache/flownets_EPE1.951.pth.tar -trainBatch 40 -sbatch True -snapshot 1 -snapshotStrategy none -gradClip 20. -imageWidth 640 -imageHeight 192

For a more detailed explanation of parameters, refer to args.py.