DenseMapNet

Keras code of "Fast Disparity Estimation using Dense Networks" paper at the International Conference on Robotics and Automation, Australia, 2018 (ICRA 2018)

DenseMapNet Features

Predicts disparity map using full resolution stereo RGB
Fast at >=30Hz on NVIDIA 1080Ti GPU
Tiny network with only 290k parameters
Accurate with Low End-Point-Error or EPE

Sample Predictions

Driving, Monkaa, and Flying Datasets

KITTI 2015

Demo

Dataset

Download datasets:

Copy: cp driving.tar.bz2 densemapnet/dataset

Change dir and extract: cd densemanpnet/dataset; tar jxvf driving.tar.bz2

Available datasets:

driving - Driving
mpi - MPI Sintel

Additional datasets will be available in the future.

Training

In some datasets, the train data is split into multiple files. For example, driving is split into 4 files while mpi fits into 1 file.

To train the network:

python3 predictor.py --dataset=driving --num_dataset=4

Alterntaively, load the pre-trained weigths:

python3 predictor.py --dataset=driving --num_dataset=4 --weights=checkpoint/driving.densemapnet.weights.h5

Testing

To measure EPE using test set:

python3 predictor.py --dataset=driving --num_dataset=4 --weights=checkpoint/driving.densemapnet.weights.h5 --notrain

To benchmark speed only:

python3 predictor.py --dataset=driving --num_dataset=4 --weights=checkpoint/driving.densemapnet.weights.h5 --predict

To generate disparity predictions on both train and test datasets (complete sequential images used to create the video):

python3 predictor.py --dataset=driving --num_dataset=4 --weights=checkpoint/driving.densemapnet.weights.h5 --predict --images

Citation

If you find this work useful, please cite:

@conference{Atienza18,
  title = {Fast Disparity Estimation using Dense Networks},
  author = {Atienza, Rowel},
  booktitle = {Proceedings 2018 IEEE International Conference on Robotics and Automation (ICRA)},
  publisher = {IEEE},
  address = {Piscataway, NJ, USA},
  month = May,
  year = {2018},
  month_numeric = {5}
}

USTCPCS / densemapnet