IndoorDepth

This is a pytorch implementation of IndoorDepth for Deeper into Self-Supervised Monocular Indoor Depth Estimation. The training codes and pretrained models are available here. It is worth noting that the training codes only support 3 frames of training. If you need to use 5 frames of images as input for training, you need to modify the content in our trainer.py. (For example, from line 312 to line 323).

Preparation

Installation

Assuming a fresh Anaconda distribution (with python 3.6), you can Install pytorch first by running

conda install pytorch=1.5.1 torchvision=0.6.1  cudatoolkit=10.1 -c pytorch

Then install other requirements

pip install -r requirements.txt

Datasets & Preprocessing

Please download preprocessed (sampled in 5 frames) NYU-Depth-V2 dataset by Junjie Hu and extract it.

Extract the superpixels and line segments by excuting. It's worth noting that line segmentation only requires the installation of any version of opencv-python lower than 3.4.6, i.e. it doesn't have to be version 3.1.0.4. The following steps can be skipped if only testing our pretrained models.

python preprocess/extract_superpixel.py --data_path data/
pip uninstall opencv_python==4.4.0.46
pip install opencv_python==3.1.0.4
python preprocess/extract_lineseg.py --data_path data/ 
pip uninstall opencv_python==3.1.0.4
pip install opencv_python==4.4.0.46

Try an image

python inference_single_image.py --image_path $IMAGE_PATH --load_weights_folder $MODEL_PATH

Training

python train.py --data_path $DATA_PATH --frame_ids 0 -2 2

Evaluation

We provide pretrained models on NYUv2 datasets. You need to uncompress it and put it into the model folder.

IndoorDepth model:

NYU Depth Estimation

python evaluate_nyu_depth.py --data_path $DATA_PATH --load_weights_folder $MODEL_PATH 

ScanNet Depth Estimation

python evaluate_scannet_depth.py --data_path $DATA_PATH --load_weights_folder $MODEL_PATH 

Note: to evaluate on ScanNet, one has to download the preprocessed data by P^2Net.

Acknowledgements

The project borrows codes from Monodepth2 , P^2Net and PLNet. Many thanks to their authors.

Citation

Please cite our papers if you find our work useful in your research.

@misc{fan2023deeper,
      title={Deeper into Self-Supervised Monocular Indoor Depth Estimation}, 
      author={Chao Fan and Zhenyu Yin and Yue Li and Feiqing Zhang},
      year={2023},
      eprint={2312.01283},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}