DilatedCRF

Pytorch implementation for fully-learnable DilatedCRF.

If you find my work helpful, please consider our paper:

@article{Mo2022dilatedcrf,
    title={Dilated Continuous Random Field for Semantic Segmentation},  
    author={Xi Mo, Xiangyu Chen, Cuncong Zhong, Rui Li, Kaidong Li, Sajid Usman},
    booktitle={IEEE International Conference on Robotics and Automation}, 
    year={2022}  
}

Easy Setup

Please install these required packages by official guidance:

python >= 3.6
pytorch >= 1.0.0
torchvision
pillow
numpy

How to Use

1. Prepare dataset

• Dowload suction-based-grasping-dataset.zip (1.6GB) [link]. Please cite relevant paper:

@article{zeng2018robotic, 
    title={Robotic Pick-and-Place of Novel Objects in Clutter with Multi-Affordance Grasping and Cross-Domain Image Matching},  
    author={Zeng, Andy and Song, Shuran and Yu, Kuan-Ting and Donlon, Elliott and Hogan, Francois Robert and Bauza, Maria and Ma, Daolin and Taylor, Orion and Liu,     Melody and Romo, Eudald and Fazeli, Nima and Alet, Ferran and Dafle, Nikhil Chavan and Holladay, Rachel and Morona, Isabella and Nair, Prem Qu and Green, Druck and Taylor, Ian and Liu, Weber and Funkhouser, Thomas and Rodriguez, Alberto},  
    booktitle={Proceedings of the IEEE International Conference on Robotics and Automation}, 
    year={2018}  
}

• Train your own semantic segmentation classifers on the suction dataset, generate training samples and test samples for DilatedCRF. You can also download my training set and test set (872MB) [link], extract the default folder dataset to the main directory.
  NOTE: Customized training and test samples must be organized the same as the default dataset format.

2. Train network

• If you want to customize training process, modify utils/configuration.py parameters according to its instructions.

• Train DilatedCRF use default dataset folder, or customized dataset path by -d argument.
  NOTE: checkpoints will be written to the default folder checkpoint.

  python DialatedCRF.py -train
  

  or restore training using the lattest .pt file stored in default folder checkpoint:

  python DialatedCRF.py -train -r
  

  or you may want to use specified checkpoint:

  python DialatedCRF.py -train -r -c path/to/your/ckpt
  

  Note that checkpoint file must match the parameter "SCALE" specified in utils/configuration.py. To specify customized dataset folder, use:

  python RGANet.py -train -d your/dataset/path
  

3. Validation

• Complete dataset folder mentioned above and a valid checkpoint are required. You can download my checkpoint for "SCALE" = 0.25 (42.4MB) [link], be sure to adjust corresponding configurations beforehand. Then run:

  python DialatedCRF.py -v
  

  or you may specify dataset folder by -d:

  python DialatedCRF.py -v -d your/path/to/dataset/folder
  

• Final results will be written to folder results. Metrics including Jaccard, F1-score, accuracy, etc., will be gathered as evaluation.txt in the folder results/evaluation

Contributed by Xi Mo,
License: Apache 2.0