Awesome

MUSTNet

the source code for MUSTNet

Useage

Install

you can install all the dependencies with:

conda install pytorch=1.2.0 torchvision=0.6.0 -c pytorch
conda install opencv=4.2
pip install scipy=1.4.1

Dataset

KITTI

The KITTI (raw) dataset used in our experiments can be downloaded from the KITTI website.

Training

[One stage] Self-supervised manner

Firstly, you can train our model in an fully self-supervised manner with

python train.py --training_stage 1 --set_assistant_teacher False

Then, you can train the model with only one coefficient decoder.

[Two stages] The first training stage for the teachers

python train.py --training_stage 1 --num_teacher 4 --set_assistant_teacher True

Then, you can train the model with 4 coefficient decoder.

KITTI Evaluation

Firstly, prepare the ground truth depth maps by running:

python export_gt_depth.py --data_path ./kitti_RAW

Then put the pretrained Models in ./models

You can evaluate the pretrained models by running:

python evaluate_depth.py --data_path ./kitti_RAW --load_weights_folder ./models/MUSTNet_S_640x192 --MUSTNet
python evaluate_depth.py --data_path ./kitti_RAW --load_weights_folder ./models/MUSTNet_S_124x320 --MUSTNet

Pretrained Models

We provided pretrained model as follow:

Precomputed Depth Maps

We also provide pre-computed depth maps for supervised training and evaluation:

References