Learning Depth from Monocular Videos using Direct Methods

Implementation of the methods in "Learning Depth from Monocular Videos using Direct Methods". If you find this code useful, please cite our paper:

@InProceedings{Wang_2018_CVPR,
author = {Wang, Chaoyang and Miguel Buenaposada, José and Zhu, Rui and Lucey, Simon},
title = {Learning Depth From Monocular Videos Using Direct Methods},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2018}
}

Dependencies

Python 3.6
PyTorch 0.3.1 (latter or eariler version of Pytorch is non-compatible.)
visdom, dominate

Training

data preparation

We refer "SfMLeaner" to prepare the training data from KITTI. We assume the processed data is put in directory "./data_kitti/".

training with different pose prediction modules

Start visdom server before for inspecting learning progress before starting the training process.

python -m visdom.server -port 8009

train from scratch with PoseNet

bash run_train_posenet.sh

see run_train_posenet.sh for details.

finetune with DDVO

Use pretrained posenet to give initialization for DDVO. Corresponds to the results reported as "PoseNet+DDVO" in the paper.

bash run_train_finetune.sh

see run_train_finetune.sh for details.

Testing

Pretrained depth network reported as "Posenet-DDVO(CS+K)" in the paper [download].
Depth prediction results on KITTI eigen test split(see Table 1 in the paper): [Posenet(K)], [DDVO(K)], [Posenet+DDVO(K)],[Posenet+DDVO(CS+K)]
To test yourself:

CUDA_VISIBLE_DEVICES=0 nice -10 python src/testKITTI.py --dataset_root $DATAROOT --ckpt_file $CKPT --output_path $OUTPUT --test_file_list test_files_eigen.txt

Evaluation

We again refer to "SfMLeaner" for their evaluation code.

Acknowledgement

Part of the code structure is borrowed from "Pytorch CycleGAN"

snooble/LKVOLearner