Home

Awesome

TODE-Trans: Transparent Object Depth Estimation with Transformer

[Paper]

PyTorch implementation of paper "TODE-Trans: Transparent Object Depth Estimation with Transformer"

Dataset Preparation

ClearGrasp Dataset

ClearGrasp can be downloaded at their official website (Both training and testing dataset are needed). After you download zip files and unzip them on your local machine, the folder structure should be like

${DATASET_ROOT_DIR}
├── cleargrasp
│   ├── cleargrasp-dataset-train
│   ├── cleargrasp-dataset-test-val

Omniverse Object Dataset

Omniverse Object Dataset can be downloaded here. After you download zip files and unzip them on your local machine, the folder structure should be like

${DATASET_ROOT_DIR}
├── omniverse
│   ├── train
│   │	├── 20200904
│   │	├── 20200910

TransCG Dataset

TransCG dataset is now available on official page.

Requirements

The code has been tested under

System dependencies can be installed by:

sudo apt-get install libhdf5-10 libhdf5-serial-dev libhdf5-dev libhdf5-cpp-11
sudo apt install libopenexr-dev zlib1g-dev openexr

Other dependencies can be installed by

pip install -r requirements.txt

Testing

We provide transcg pretrained checkpoints at checkpoints/.

Training

#Train on transcg dataset and test on transcg
python train.py -c ./configs/train_transcg_val_transcg.yaml

#Tran on CGsyn+ood and test on CGreal
python train.py -c ./configs/train_cgsyn+ood_val_cgreal.yaml
#Tran on CGsyn+ood and test on Transcg
python train.py -c ./configs/train_cgsyn+ood_val_transcg.yaml

Citation

@article{2022tode,
    title   = {TODE-Trans: Transparent Object Depth Estimation with Transformer},
    author  = {Kang Chen, Shaochen Wang, Beihao Xia, Dongxu Li, Zhen Kan, and Bin Li},
    journal = {arXiv preprint arXiv:2209.08455}
    year    = {2022}
}