Awesome

Pytorch Material Classification

This repo provides examples for material classification in GTOS, GTOS-MOBILE, DTD and MINC dataset using PyTorch.

Setup

Prerequisites

Ubuntu
Pytorch
- pip3 install torch torchvision
Easydict
- pip3 install easydict
tqdm
- pip3 install tqdm
Pytorch-Encoding
- pip3 install torch-encoding
- Note: You need to install Pytorch 1.0 for torch-encoding, or you can modify the encoding layer based on this for latest Pytorch.

Getting Started

Clone this repo:

git clone https://github.com/jiaxue1993/pytorch-material-classification.git
cd pytorch-material-classification/

Download GTOS, GTOS_MOBILE, DTD, MINC to the dataset folder
Navigate to different experiment folder and train network. For example, you can finetune ResNet on GTOS-MOBILE dataset with followint command

cd experiments/gtos_mobile.finetune.resnet/
python train.py

Accuracy & Statistics

<table class="tg"> <thead> <tr> <th class="tg-5unb">Base Model</th> <th class="tg-5unb">Dataset</th> <th class="tg-5unb">Method</th> <th class="tg-5unb">Accuracy</th> <th class="tg-5unb">Pretrained Model<br></th> </tr> </thead> <tbody> <tr> <td class="tg-lcl0" rowspan="4">MobileNet</td> <td class="tg-wp8o" rowspan="4"><span style="font-weight:normal;font-style:normal;text-decoration:none">GTOS</span></td> <td class="tg-wp8o">Finetune</td> <td class="tg-wp8o">80.4</td> <td class="tg-wp8o" rowspan="6"><a href="https://1drv.ms/u/s!AmTf4gl42Obncc3EohCJubNVHZQ?e=PZ1eAC" target="_blank" rel="noopener noreferrer">One Drive</a></td> </tr> <tr> <td class="tg-wp8o">DAIN</td> <td class="tg-wp8o">82.5</td> </tr> <tr> <td class="tg-wp8o">DEP</td> <td class="tg-wp8o">83.3</td> </tr> <tr> <td class="tg-wp8o">TEAN</td> <td class="tg-wp8o">84.7</td> </tr> <tr> <td class="tg-lcl0" rowspan="2">ResNet-50</td> <td class="tg-wp8o">DTD</td> <td class="tg-wp8o"><span style="font-weight:normal;font-style:normal;text-decoration:none">DEP</span></td> <td class="tg-wp8o">73.2</td> </tr> <tr> <td class="tg-wp8o">MINC</td> <td class="tg-wp8o"><span style="font-weight:normal;font-style:normal;text-decoration:none">DEP</span></td> <td class="tg-wp8o">82.0</td> </tr> </tbody> </table>

Citation

Please consider citing following projects in your publications if it helps your research.

Differential Angular Imaging for Material Recognition [pdf]

@inproceedings{xue2017differential,
  title={Differential angular imaging for material recognition},
  author={Xue, Jia and Zhang, Hang and Dana, Kristin and Nishino, Ko},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={764--773},
  year={2017}
}

Deep Texture Manifold for Ground Terrain Recognition [pdf]

@inproceedings{xue2018deep,
  title={Deep texture manifold for ground terrain recognition},
  author={Xue, Jia and Zhang, Hang and Dana, Kristin},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={558--567},
  year={2018}
}

Deep TEN: Texture Encoding Network [pdf]

@inproceedings{zhang2017deep,
  title={Deep ten: Texture encoding network},
  author={Zhang, Hang and Xue, Jia and Dana, Kristin},
  booktitle={Proceedings of the IEEE conference on computer vision and pattern recognition},
  pages={708--717},
  year={2017}
}

Differential Viewpoints for Ground Terrain Material Recognition [pdf][arxiv]

@article{xue2020differential,
  title={Differential Viewpoints for Ground Terrain Material Recognition},
  author={Xue, Jia and Zhang, Hang and Nishino, Ko and Dana, Kristin},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2020},
  publisher={IEEE}
}

Acknowledgement

Part of the code comes from PyTorch-Encoding, TorchSeg, pytorch-mobilenet-v2