Skip to content

[WACV2023] Frequency Aware Self-supervised Depth Estimation

License

Notifications You must be signed in to change notification settings

xingyuuchen/freq-aware-depth

Repository files navigation

Freq-Aware-Depth

This is the official PyTorch implementation of our paper "Frequency-Aware Self-Supervised Depth Estimation" (WACV 2023)

We introduce FreqAwareDepth, with highly generalizable performance-boosting features that can be easily integrated into other models, see the paper for more details.

Our methods introduce no more than 10% extra training time and no extra inference time at all.

🎬 Watch our video.

🍻 KITTI Results

generalizability

🛠️ Setup

Assuming a fresh Anaconda distribution, you can install the dependencies with:

conda install pytorch=1.7.1 torchvision=0.8.2 -c pytorch 
pip install tensorboardX==1.5     # 1.4 also ok
conda install opencv=3.4.2    # just needed for evaluation, 3.3.1 also ok

Our code is build upon monodepth2.

🚄 Training

Train our full model:

python train.py --model_name {name_you_expect}

KITTI training data

You can download the entire KITTI RAW dataset by running:

wget -i splits/kitti_archives_to_download.txt -P kitti_data/

Then unzip with

cd kitti_data
unzip "*.zip"
cd ..

Warning: it weighs about 175GB.

👀 Predict depth for a single image

python test_simple.py --image_path assets/test_image.jpg --model_path {pretrained_model_path}

💯 KITTI evaluation

To evaluate, run:

python evaluate_depth.py --eval_mono --load_weights_folder {model_path}

🐷 Note: Make sure you have run the command below to generate ground truth depth before evaluating.

python export_gt_depth.py --data_path {KITTI_path} --split eigen

✏️ Citation

If you find our work useful or interesting, please consider citing our paper:

@inproceedings{chen2023frequency,
  title={{Frequency-Aware Self-Supervised Monocular Depth Estimation}},
  author={Chen, Xingyu and Li, Thomas H and Zhang, Ruonan and Li, Ge},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={5808--5817},
  year={2023}
}