DiPE

The Pytorch code for our following papers

DiPE: Deeper into Photometric Errors for Unsupervised Learning of Depth and Ego-motion from Monocular Videos, IROS 2020 (pdf)

Unsupervised Monocular Depth Perception: Focusing on Moving Objects, accepted by IEEE Sensors Journal

Hualie Jiang, Laiyan Ding, Zhenglong Sun and Rui Huang

Preparation

Installation

Install pytorch first by running

conda install pytorch=1.0.0 torchvision=0.2.1  cuda100 -c pytorch

Then install other requirements

pip install -r requirements.txt

Datasets

Please download and preprocess the KITTI dataset as Monodepth2 does.

For the Cityscapes dataset, one should download gtFine_trainvaltest.zip, leftimg8bit_sequence_trainvaltest.zip, and disparity_trainvaltest.zip and extract them to one directory.

Training

Monocular Depth

KITTI Eigen Zhou Split

python train.py --data_path $DATA_PATH(raw_data) --model_name dipe_eigen --split kitti_eigen_zhou --dataset kitti 

KITTI Official Benchmark Split

python train.py --data_path $DATA_PATH(raw_data) --model_name dipe_bench --split kitti_benchmark --dataset kitti 

Unsupervised Monocular Depth Perception: Focusing on Moving ObjectsCityscapes

python train.py --data_path $DATA_PATH(cityscapes) --model_name cityscapes --split cityscapes --dataset cityscapes --png --frame_ids 0 -2 2 

Monocular Odometry

Pair-frames

python train.py --data_path $DATA_PATH(odometry) --model_name dipe_odom2 --split kitti_odom --dataset kitti_odom \ 
--frame_ids 0 -1 1 --pose_model_input pairs

3-frames

python train.py --data_path $DATA_PATH(odometry) --model_name dipe_odom3 --split kitti_odom --dataset kitti_odom \
--frame_ids 0 -1 1 --pose_model_input all

5-frames

python train.py --data_path $DATA_PATH(odometry) --model_name dipe_odom5 --split kitti_odom --dataset kitti_odom \
--frame_ids 0 -2 -1 1 2 --pose_model_input all --disable_occlusion_mask_from_photometric_error

Evaluation

The pretrained models of our paper is available on GoogleDrive and BaiduDisk (code:rq3a).

Monocular Depth

KITTI Eigen Split

python evaluate_kitti.py --data_path $DATA_PATH(raw_data) --load_weights_folder $MODEL_PATH(dipe_eigen) \ 
--eval_mono --eval_split kitti_eigen

KITTI Official Benchmark Split

python evaluate_kitti.py --data_path $DATA_PATH(depth) --load_weights_folder $MODEL_PATH(dipe_bench) --dataset kitti_depth \ 
--eval_mono --eval_split kitti_benchmark

Results

Split	Abs Rel	Sq Rel	RMSE	RMSE_log	a1	a2	a3
Eigen	0.112	0.875	4.795	0.190	0.880	0.960	0.981
Benchmark	0.086	0.556	3.923	0.133	0.928	0.983	0.994

KITTI Eigen Split with Background and Moving Objects Separately

python train.py  evaluate_kitti_eigen_moving_objects.py --data_path $DATA_PATH(raw_data) --load_weights_folder $MODEL_PATH(dipe_eigen) \ 
--eval_mono --eval_split kitti_eigen

Results

Region	Abs Rel	Sq Rel	RMSE	RMSE_log	a1	a2	a3
Background	0.107	0.784	4.614	0.180	0.886	0.964	0.984
Dynamic Objects	0.215	3.083	7.172	0.319	0.737	0.883	0.931

CityScapes with Background and Objects Separately

python train.py  evaluate_cityscapes.py --data_path $DATA_PATH(cityscapes) --load_weights_folder $MODEL_PATH(dipe_cityscapes) --eval_mono

Results

Region	Abs Rel	Sq Rel	RMSE	RMSE_log	a1	a2	a3
Background	0.155	2.381	8.127	0.220	0.808	0.947	0.980
Objects	0.365	13.401	9.742	0.336	0.697	0.861	0.924

Monocular Odometry

Divide the poses for different tests

python ./splits/odom/kitti_divide_poses.py --data_path $DATA_PATH(odometry) 

2-pair-frames

python evaluate_kitti.py --data_path $DATA_PATH(odometry) --load_weights_folder $MODEL_PATH(dipe_odom2) \ 
--dataset kitti_odom --pose_model_input pairs --eval_split kitti_odom_09 

3-all-frames

python evaluate_kitti.py --data_path $DATA_PATH(odometry) --load_weights_folder $MODEL_PATH(dipe_odom3) \ 
--dataset kitti_odom --frame_ids 0 -1 1 --pose_model_input all --eval_split kitti_odom_09 

5-all-frames

python evaluate_kitti.py --data_path $DATA_PATH(odometry) --load_weights_folder $MODEL_PATH(dipe_odom5) \
--dataset kitti_odom --frame_ids 0 -2 -1 1 2 --pose_model_input all --eval_split kitti_odom_09 

To test sequence 10, we need to set --eval_split odom_10.

Results

#frames	Sequence 09		Sequence 10
	ATE	RE	ATE	RE
2	0.0125 ± 0.0055	0.0023 ± 0.0011	0.0122 ± 0.0081	0.0027 ± 0.0019
3	0.0122 ± 0.0057	0.0025 ± 0.0012	0.0120 ± 0.0082	0.0029 ± 0.0020
5	0.0120 ± 0.0057	0.0026 ± 0.0014	0.0118 ± 0.0082	0.0030 ± 0.0022

Acknowledgements

The project is built upon Monodepth2. We thank Monodepth2's authors for their excellent work and repository.

Citation

Please cite our papers if you find our work useful in your research.

@inproceedings{jiang2020dipe,
  title={DiPE: Deeper into Photometric Errors for Unsupervised Learning of Depth and Ego-motion from Monocular Videos},
  author={Jiang, Hualie and Ding, Laiyan and Sun, Zhenglong and Huang, Rui},
  booktitle={In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year={2020}
}
@article{jiang2021unsupervised,
	  title={Unsupervised Monocular Depth Perception: Focusing on Moving Objects},
	  author={Jiang, Hualie and Ding, Laiyan and Sun, Zhenglong and Huang, Rui},
	  journal={IEEE Sensors Journal},
	  year={2021},
	  publisher={IEEE}
}