TOM-Net

TOM-Net: Learning Transparent Object Matting from a Single Image, CVPR 2018 (Spotlight),
Guanying Chen*, Kai Han*, Kwan-Yee K. Wong
(* equal contribution)

This paper addresses the problem of transparent object matting from a single image.

Dependencies

TOM-Net is implemented in Torch and tested with Ubuntu 14.04. Please install Torch first following the official document.

• python 2.7
• numpy
• cv2
• CUDA-8.0
• CUDNN v5.1
• Torch STN (qassemoquab/stnbhwd)

  # Basic installation steps for stn
  git clone https://github.com/qassemoquab/stnbhwd.git
  cd stnbhwd
  luarocks make

Overview

We provide:

• Pretrained model
• Datasets: Train (40GB), Validation (196MB), Test (179MB)
• Code to test model on new images
• Evaluation code on both the validation and testing data
• Instructions to train the model
• Example code for synthetic data rendering
• Code and models used in the journal extension (New!)

If the automatic downloading scripts are not working, please download the trained models and the introduced dataset from BaiduYun (Models and Datasets).

Testing

Download Pretrained Model

sh scripts/download_pretrained_model.sh

If the above command is not working, please manually download the trained models from BaiduYun (Models, Datasets) and put them in ./data/models/.

Test on New Images

# Replace ${gpu} with the selected GPU ID (starting from 0)

# Test a single image without having the background image
CUDA_VISIBLE_DEVICES=${gpu} th eval/run_model.lua -input_img images/bull.jpg 

# You can find the results in data/TOM-Net_model/

Evaluation on Synthetic Validation Data

# Download synthetic validation dataset
sh scripts/download_validation_dataset.sh

# Quantitatively evaluate TOM-Net on different categories of synthetic object 
# Replace ${class} with one of the four object categories (glass, water, lens, cplx)
CUDA_VISIBLE_DEVICES=${gpu} th eval/run_synth_data.lua -img_list ${class}.txt

# Similarly, you can find the results in data/TOM-Net_model/

Evaluation on Real Testing Data

# Download real testing dataset, 
sh scripts/download_testing_dataset.sh

# Test on sample images used in the paper
CUDA_VISIBLE_DEVICES=${gpu} th eval/run_model.lua -img_list Sample_paper.txt

# Quantitatively evaluate TOM-Net on different categories of real-world object 
# Replace ${class} with one of the four object categories (Glass, Water, Lens, Cplx)
CUDA_VISIBLE_DEVICES=${gpu} th eval/run_model.lua -img_list ${class}.txt  

Training

To train a new TOM-Net model, please follow the following steps:

• Download the training data

# The size of the zipped training dataset is 40 GB and you need about 207 GB to unzip it.
sh scripts/download_training_dataset.sh

• Train CoarseNet on simple objects

CUDA_VISIBLE_DEVICES=$gpu th main.lua -train_list train_simple_98k.txt -nEpochs 13 -prefix 'simple'
# Please refer to opt.lua for more information about the training options

# You can find log file, checkpoints and visualization results in data/training/simple_*

• Train CoarseNet on both simple and complex objects

# Finetune CoarseNet with all of the data
CUDA_VISIBLE_DEVICES=$gpu th main.lua -train_list train_all_178k.txt -nEpochs 7 -prefix 'all' -retrain data/training/simple_*/checkpointdir/checkpoint13.t7

# You can find log file, checkpoints and visualization results in data/training/all_*

• Train RefineNet on both simple and complex objects

CUDA_VISIBLE_DEVICES=$gpu th refine/main_refine.lua -train_list train_all_178k.txt -nEpochs 20 -coarse_net data/training/all_*/checkpointdir/checkpoint7.t7 
# Train RefineNet with all of the data
# Please refer to refine/opt_refine.lua for more information about the training options

# You can find log file, checkpoints and visualization results in data/training/all_*/refinement/

Synthetic Data Rendering

Please refer to TOM-Net_Rendering for sample rendering codes.

Codes and Models Used in the Journal Extension (IJCV)

Test TOM-Net^+Bg and TOM-Net^+Trimap on Sample Images

# Download pretrained models
sh scripts/download_pretrained_models_IJCV.sh

# Test TOM-Net+Bg on sample images
CUDA_VISIBLE_DEVICES=${gpu} th eval/run_model.lua -input_root images/TOM-Net_with_Trimap_Bg_Samples/ -img_list img_bg_trimap_list.txt -in_bg -c_net data/TOM-Net_plus_Bg_Model/CoarseNet_plus_Bg.t7 -r_net data/TOM-Net_plus_Bg_Model/RefineNet_plus_Bg.t7 
# You can find the results in data/TOM-Net_plus_Bg_Model/*

# Test TOM-Net+Trimap on sample images
CUDA_VISIBLE_DEVICES=${gpu} th eval/run_model.lua -input_root images/TOM-Net_with_Trimap_Bg_Samples/ -img_list img_bg_trimap_list.txt -in_trimap -c_net data/TOM-Net_plus_Trimap_Model/CoarseNet_plus_Trimap.t7 -r_net data/TOM-Net_plus_Trimap_Model/RefineNet_plus_Trimap.t7 
# You can find the results in data/TOM-Net_plus_Trimap_Model/*

Train TOM-Net^+Bg and TOM-Net^+Trimap

To train a new TOM-Net^+Bg or TOM-Net^+Trimap model, please follow the same procedures as training TOM-Net, except that you need to append -in_bg or -in_trimap at the end of the commands.

Citation

If you find this code or the provided data useful in your research, please consider cite the following relevant paper(s):

@inproceedings{chen2018tomnet,
  title={TOM-Net: Learning Transparent Object Matting from a Single Image},
  author={Chen, Guanying and Han, Kai and Wong, Kwan-Yee~K.},
  booktitle={CVPR},
  year={2018}
}

@inproceedings{chen2019LTOM,
  title={Learning Transparent Object Matting},
  author={Chen, Guanying and Han, Kai and Wong, Kwan-Yee~K.},
  booktitle={IJCV},
  year={2019}
}