This code provides a combination of DSO and Monodepth. For every keyframe, depth values are initialized with the prediction from Monodepth.
Absolute keyframe trajectory RMSE (in meter) on KITTI dataset (DSO and ORB-SLAM numbers are from CNN-SVO paper)
Sequence on KITTI | CNN-DSO | DSO | ORB-SLAM |
---|---|---|---|
00 | 15.13 | 113.18 | 77.95 |
01 | 5.901 | X | X |
02 | 12.53 | 116.81 | 41.00 |
03 | 1.516 | 1.3943 | 1.018 |
04 | 0.100 | 0.422 | 0.930 |
05 | 20.3 | 47.46 | 40.35 |
06 | 1.547 | 55.61 | 52.22 |
07 | 8.369 | 16.71 | 16.54 |
08 | 10.53 | 111.08 | 51.62 |
09 | 14.00 | 52.22 | 58.17 |
10 | 4.10 | 11.09 | 18.47 |
- Setup dependencies of DSO (https://github.com/JakobEngel/dso)
- Build TensorFlow C++ API (https://github.com/yan99033/monodepth-cpp/tree/master/Tensorflow_build_instructions). This is the hardest part!
- Build monodepth-cpp (https://github.com/yan99033/monodepth-cpp).
- Prepare Monodepth pre-trained model. You can freaze .ckpt or download the model trained on cityscapes and fine-tuned on kitti here.
-
Download the repository.
git clone https://github.com/muskie82/CNN-DSO.git
-
Modify paths to include directories and libraries of TensorFlow and monodepth-cpp in
CMakeLists.txt
(4 lines of/abosolute/path/to/XXXXX
). -
Build
cd CNN-DSO mkdir build cd build cmake .. make -j4
In addition to original DSO command line, you should specify the path to pre-trained model by cnn
.
bin/dso_dataset \
files=XXXXX/sequence_XX/image_0 \
calib=XXXXX/sequence_XX/camera.txt \
cnn=XXXXX/model_city2kitti.pb \
preset=0 \
mode=1
- Direct Sparse Odometry, Engel, Jakob, Vladlen Koltun, and Daniel Cremers, IEEE transactions on pattern analysis and machine intelligence 40.3 (2018): 611-625. (https://github.com/JakobEngel/dso)
- Unsupervised monocular depth estimation with left-right consistency, *Godard, Clément, Oisin Mac Aodha, and Gabriel J. Brostow. *, CVPR. Vol. 2. No. 6. 2017. (https://github.com/mrharicot/monodepth)
- CNN-SVO: Improving the Mapping in Semi-Direct Visual Odometry Using Single-Image Depth Prediction., Loo, S. Y., Amiri, A. J., Mashohor, S., Tang, S. H., and Zhang, H, arXiv preprint arXiv:1810.01011 (2018). (https://github.com/yan99033/CNN-SVO)
- stereo_dso by HorizonAD: https://github.com/HorizonAD/stereo_dso
GPLv3 license. I don't take any credit from DSO, Monodepth and monodepth-cpp. Please check them.