SC-SfMLearner

This codebase implements the system described in the paper:

Unsupervised Scale-consistent Depth and Ego-motion Learning from Monocular Video

Jia-Wang Bian, Zhichao Li, Naiyan Wang, Huangying Zhan, Chunhua Shen, Ming-Ming Cheng, Ian Reid

NeurIPS 2019 [PDF] [Project webpage]

Point cloud visulization on KITTI (left) and real-world data (right)

Dense reconstruction (left) using the estimated depth (bottom right)

Core contributions

1. A geometry consistency loss, which makes the predicted depths to be globally scale consistent.
2. A self-discovered mask, which detects moving objects and occlusions effectively and efficiently.
3. The scale-consistent predictions allow for doing Monocular Visual Odometry on long videos.

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

@inproceedings{bian2019depth,
  title={Unsupervised Scale-consistent Depth and Ego-motion Learning from Monocular Video},
  author={Bian, Jia-Wang and Li, Zhichao and Wang, Naiyan and Zhan, Huangying and Shen, Chunhua and Cheng, Ming-Ming and Reid, Ian},
  booktitle= {Thirty-third Conference on Neural Information Processing Systems (NeurIPS)},
  year={2019}
}

Updates (Compared with NeurIPS version)

Note that this is an updated version, and you can find the original version in 'Release / NeurIPS Version' for reproducing the results reported in paper. Compared with NeurIPS version, we (1) Change networks by using Resnet18 and Resnet50 pretrained model (on ImageNet) for depth and pose encoders. (2) We add 'auto_mask' by Monodepth2 to remove stationary points.

We add training and testing on NYUv2 indoor depth dataset. See Unsupervised-Indoor-Depth for details.

Preamble

This codebase was developed and tested with python 3.6, Pytorch 1.0.1, and CUDA 10.0 on Ubuntu 16.04. It is based on Clement Pinard's SfMLearner implementation.

Prerequisite

pip3 install -r requirements.txt

or install manually the following packages :

torch >= 1.5.1
imageio
matplotlib
scipy
argparse
tensorboardX
blessings
progressbar2
path

It is also advised to have python3 bindings for opencv for tensorboard visualizations

Datasets

See "scripts/run_prepare_data.sh".

For KITTI Raw dataset, download the dataset using this script http://www.cvlibs.net/download.php?file=raw_data_downloader.zip) provided on the official website.

For KITTI Odometry dataset, download the dataset with color images.

Or you can download our pre-processed dataset from the following link

kitti_256 (for kitti raw) | kitti_vo_256 (for kitti odom) | kitti_depth_test (eigen split) | kitti_vo_test (seqs 09-10)

Training

The "scripts" folder provides several examples for training and testing.

You can train the depth model on KITTI Raw by running

sh scripts/train_resnet18_depth_256.sh

or train the pose model on KITTI Odometry by running

sh scripts/train_resnet50_pose_256.sh

Then you can start a tensorboard session in this folder by

tensorboard --logdir=checkpoints/

and visualize the training progress by opening https://localhost:6006 on your browser.

Evaluation

You can evaluate depth on Eigen's split by running

sh scripts/test_kitti_depth.sh

evaluate visual odometry by running

sh scripts/test_kitti_vo.sh

and visualize depth by running

sh scripts/run_inference.sh

Pretrained Models

Latest Models

To evaluate the NeurIPS models, please download the code from 'Release/NeurIPS version'.

Depth Results (Updated version, KITTI raw dataset, using the Eigen's splits)

Visual Odometry Results (Updated version, KITTI odometry dataset, trained on 00-08)

Related projects

• SfMLearner-Pytorch (CVPR 2017, our baseline framework.)

• Depth-VO-Feat (CVPR 2018, trained on stereo videos for depth and visual odometry)

• DF-VO (ICRA 2020, use scale-consistent depth with optical flow for more accurate visual odometry)

• Kitti-Odom-Eval-Python (python code for kitti odometry evaluation)

• Unsupervised-Indoor-Depth (Using SC-SfMLearner in NYUv2 dataset)