GeoNet

This is a Tensorflow implementation of our paper:

GeoNet: Unsupervised Learning of Dense Depth, Optical Flow and Camera Pose (CVPR 2018)

Zhichao Yin and Jianping Shi

arxiv preprint: (https://arxiv.org/abs/1803.02276)

Requirements

This code has been tested with Python 2.7, TensorFlow 1.1 and CUDA 8.0 on Ubuntu 16.04.

Data preparation

For replicating our results in all of the three tasks (monocular depth, camera pose and optical flow), you need to download the following datasets, and preprocess them into certain formats:

KITTI

For depth and flow tasks, the training data is KITTI raw dataset and you can download them by the official script;

For pose task, the training data is KITTI odometry dataset and you should download the calibration files as well as ground truth poses (for evaluation).

After downloaded the data, you can run the following command for preprocessing:

python data/prepare_train_data.py --dataset_dir=/path/to/kitti/dataset/ --dataset_name=kitti_split --dump_root=/path/to/formatted/data/ --seq_length=3 --img_height=128 --img_width=416 --num_threads=16 --remove_static

For depth task, the --dataset_name should be kitti_raw_eigen and --seq_length is set to 3;

For flow task, the --dataset_name should be kitti_raw_stereo and --seq_length is set to 3;

For pose task, the --dataset_name should be kitti_odom and --seq_length is set to 5.

Cityscapes

You can optionally pretrain the model on Cityscapes dataset for any of the three tasks. The required training data is image sequence leftImg8bit_sequence_trainvaltest.zip and calibration file camera_trainvaltest.zip. After downloaded them, simply run:

python data/prepare_train_data.py --dataset_dir=/path/to/cityscapes/dataset/ --dataset_name='cityscapes' --dump_root=/path/to/formatted/data/ --seq_length=3 --img_height=171 --img_width=416 --num_threads=16

Training

Our code supports two training modes, corresponding to our stage-wise training strategy. The train_rigid mode is mainly for learning depth and pose, while train_flow mode supports direct or residual flow learning.

For train_rigid mode (depth and pose tasks), run the command

python geonet_main.py --mode=train_rigid --dataset_dir=/path/to/formatted/data/ --checkpoint_dir=/path/to/save/ckpts/ --learning_rate=0.0002 --seq_length=3 --batch_size=4 --max_steps=350000 

You can switch the network encoder by setting --dispnet_encoder flag, or perform depth scale normalization (see this paper for details) by setting --scale_normalize as True. Note that for replicating depth and pose results, the --seq_length is suggested to be 3 and 5 respectively.

For train_flow mode (flow task), run the command

python geonet_main.py --mode=train_flow --dataset_dir=/path/to/formatted/data/ --checkpoint_dir=/path/to/save/ckpts/ --learning_rate=0.0002 --seq_length=3 --flownet_type=direct --max_steps=400000

You can choose to learn direct or residual flow by setting --flownet_type flag. Note that when the --flownet_type is residual, the --init_ckpt_file should be specified to point at a model pretrained on the same dataset with mode of train_rigid. Also a max_steps more than 200 epochs is preferred for learning residual flow.

Pretrained models

You can download our pretrained models as well as their predictions in all of the three tasks from [Google Drive]. Note that they were trained according to different splits of KITTI as described in the paper. Following the testing and evaluation instructions below, you should obtain similar results as reported in the paper.

Notes about depth scale normalization

Following most of the original hyperparameters but setting --scale_normalize as True, we have trained our DepthNet better on the Eigen's split of KITTI. The pretrained model is also provided, namely model_sn in geonet_depthnet subfolder. Note this is not included in our paper, but the performance is further improved:

Abs Rel	Sq Rel	RMSE	RMSE(log)	Acc.1	Acc.2	Acc.3
0.149	1.060	5.567	0.226	0.796	0.935	0.975

Testing

We provide testing and evaluation scripts for all of the three tasks.

Monocular Depth

Run the following command

python geonet_main.py --mode=test_depth --dataset_dir=/path/to/kitti/raw/dataset/ --init_ckpt_file=/path/to/trained/model/ --batch_size=1 --depth_test_split=eigen --output_dir=/path/to/save/predictions/

Then you can evaluate the prediction by running

python kitti_eval/eval_depth.py --split=eigen --kitti_dir=/path/to/kitti/raw/dataset/ --pred_file=/path/to/predictions/

Camera Pose

Firstly assuming you have downloaded the KITTI odometry dataset (including groundtruth poses), run

python geonet_main.py --mode=test_pose --dataset_dir=/path/to/kitti/odom/dataset/ --init_ckpt_file=/path/to/trained/model/ --batch_size=1 --seq_length=5 --pose_test_seq=9 --output_dir=/path/to/save/predictions/

Now you have predicted pose snippets. You can generate the groundtruth pose snippets by running

python kitti_eval/generate_pose_snippets.py --dataset_dir=/path/to/kitti/odom/dataset/ --output_dir=/path/to/save/gtruth/pose/snippets/ --seq_id=09 --seq_length=5

Then you can evaluate your predictions by

python kitti_eval/eval_pose.py --gtruth_dir=/path/to/gtruth/pose/snippets/ --pred_dir=/path/to/predicted/pose/snippets/

Optical Flow

Firstly you need to download the KITTI flow 2015 dataset and its multi-view extension. For replicating our flow results in the paper, a seq_length of 3 is recommended. You need to format the testing data by running

python kitti_eval/generate_multiview_extension.py --dataset_dir=/path/to/data_scene_flow_multiview/ --calib_dir=/path/to/data_scene_flow_calib/ --dump_root=/path/to/formatted/testdata/ --cam_id=02 --seq_length=3

Then you can test your trained model by

python geonet_main.py --mode=test_flow --dataset_dir=/path/to/formatted/testdata/ --init_ckpt_file=/path/to/trained/model/ --flownet_type=direct --batch_size=1 --output_dir=/path/to/save/predictions/

We again provide evaluation script:

python kitti_eval/eval_flow.py --dataset_dir=/path/to/kitti_stereo_2015/ --pred_dir=/path/to/predictions/

Acknowledgements

We thank Tinghui Zhou and Clément Godard for their great works and repos.

Reference

If you find our work useful in your research please consider citing our paper:

@inproceedings{yin2018geonet,
  title     = {GeoNet: Unsupervised Learning of Dense Depth, Optical Flow and Camera Pose},
  author    = {Yin, Zhichao and Shi, Jianping},
  booktitle = {CVPR},
  year = {2018}
}