PyTorch implementation of our ICCV2021 paper:

StructDepth: Leveraging the structural regularities for self-supervised indoor depth estimation

Boying Li*, Yuan Huang*, Zeyu Liu, Danping Zou, Wenxian Yu

(* Equal Contribution) Please consider citing our paper in your publications if the project helps your research.

@inproceedings{structdepth,
  title={StructDepth: Leveraging the structural regularities for self-supervised indoor depth estimation},
  author={Li, Boying and Huang, Yuan and Liu, Zeyu and Zou, Danping and Yu, Wenxian},
  booktitle={Proceedings of the IEEE International Conference on Computer Vision},
  year={2021}
}

The Python and PyTorch versions we use:

python=3.6

pytorch=1.7.1=py3.6_cuda10.1.243_cudnn7.6.3_0

Step1: Creating a virtual environment

conda create -n struct_depth python=3.6
conda activate struct_depth
conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=10.1 -c pytorch

Step2: Download the modified scikit_image package , in which the input parameters of the Felzenswalb algorithm have been changed to accommodate our method.

unzip scikit-image-0.17.2.zip
cd scikit-image-0.17.2
python setup.py build_ext -i
pip install -e .

Step3: Installing other packages

pip install -r requirements.txt

Please download pretrained models and unzip them to MODEL_PATH

python inference_single_image.py --image_path=/path/to/image --load_weights_folder=MODEL_PATH

Please download test dataset

It is recommended to unpack all test data and training data into the same data path and then modify the DATA_PATH when running a training or evaluation script.

Modify the evaluation script in eval.sh to evaluate NYUv2/InteriorNet/ScanNet depth and norm separately

python evaluation/nyuv2_eval_norm.py \
  --data_path DATA_PATH \
  --load_weights_folder MODEL_PATH \

The raw NYU dataset is about 400G and has 590 videos. You can download the raw datasets from there

python extract_vps_nyu.py --data_path DATA_PATH --output_dir VPS_PATH --failed_list TMP_LIST -- thresh 60 

If you need to train with a random flip, run the main direction extraction script on the images before and after the flip(add --flip) in advance, and note the failure examples, which can be skipped by referring to the code in datasets/nyu_datases.py.

Modify the training script train.sh for PATH or different trainning settings.

python train.py \
  --data_path DATA_PATH \
  --val_path DATA_PATH \
  --train_split ./splits/nyu_train_0_10_20_30_40_21483-exceptfailed-21465.txt \
  --vps_path VPS_PATH \
  --log_dir LOG_PATH \
  --model_name 1 \
  --batch_size 32 \
  --num_epochs 50 \
  --start_epoch 0 \
  --using_disp2seg \
  --using_normloss \
  --load_weights_folder PRETRAIN_MODEL_PATH \
  --lambda_planar_reg 0.1 \
  --lambda_norm_reg 0.05 \
  --planar_thresh 200 \

We borrowed a lot of codes from scikit-image, monodepth2, P²Net, and LEGO. Thanks for their excellent works!