ENG

ENG: End-to-end neural geometry for robust depth and pose estimation using CNNs

eng

Requirements

The code is tested on CUDA 9.0 + cudNN 7.4.2 + Python 3.5.2 + Tensorflow 1.9

Download the weights (requires wget)

./download_model.sh

Running

form the root dir of the project run:

python -m test_depth_flow_pose

Run Options

There are some configuration parameters, use the following command to print the help menu

python -m test_depth_flow_pose --help

Citing the Paper(s)

If you found this repository useful please cite the following:

@article{dharmasiri2018eng,
  title={ENG: End-to-end Neural Geometry for Robust Depth and Pose Estimation using CNNs},
  author={Dharmasiri, Thanuja and Spek, Andrew and Drummond, Tom},
  journal={arXiv preprint arXiv:1807.05705},
  year={2018}
}

Single Image Depth Prediction Results (KITTI 0-50m)

RMSE (m)	Rel_abs	Accuracy (δ)	Accuracy (δ²)	Accuracy (δ³)
3.284	0.092	90.6%	97.1%	98.9%

Single Image Depth Prediction Results (NYUv2 [using the indoor model] )

RMSE (m)	Rel_abs	Accuracy (δ)	Accuracy (δ²)	Accuracy (δ³)
0.478	0.111	87.2%	97.8%	99.5%

Pose Estimation (KITTI)

Sequence	ATE (m)	RPE (m)	RPE (°)
9	16.55	0.047	0.128
10	9.846	0.039	0.138