Dataset and CNN-based Models for Indoor Surface Classification

In this project, we generated a dataset that contains three different types of indoor floor surfaces: carpet, tile and wood. Then, we used this dataset to train eight CNN-based models, including our proposed model, MobileNetV2-modified.

Details of the Generated Dataset

---

• The dataset comprises a total of 2081 samples, consisting of images captured with cameras in various indoor environments and lighting conditions.
• These images were taken from different angles in accordance with the overall dimensions of the indoor robots.
• This dataset includes samples collected from more than 20 different indoor environments.
• The dataset consists of 870 carpet samples, 638 tile samples and 573 wood surface samples.

Dataset	Carpet	Tiles	Wood	Total
Train	698	510	457	1665
Test	86	64	58	208
Validation	86	64	58	208
Total	870	638	573	2081

Training and Testing

---

• The images in the dataset were saved in RGB format and resized to an equivalent size before being fed into CNN models.
• The dataset was split into three sets: 80% for training, 10% for validation and 10% for testing.
• Various CNN-based deep learning models, including InceptionV3, Xception, VGG16, VGG19, Resnet50, InceptionResnetV2, MobilenetV2, and MobileNetV2-Modified, were trained using this dataset.
• In the training process, seven different optimizers were employed, namely SGD, Adam, Adamax, RMSprop, Adagrad, Adadelta, and Nadam.
• Each model achieved a high level of accuracy. Notably, the MobileNetV2-Modified model, a modified version of the MobileNetV2 model, achieved the highest accuracy among all the models. Throughout the training and testing stages, we used libraries such as TensorFlow, OpenCV, Matplotlib, and NumPy.
• The codes used for training each model with the dataset can be accessed in the "Codes" folder.
• The code for testing the trained model using the ROS platform on the Kobuki robot is under the "Codes" folder.
• The weights that achieved the best performance for each trained model are available in the provided Google Drive link: link:https://drive.google.com/drive/folders/1aT0vXDsYdLxDKBqufeg2MVFUHPT9jl1W?usp=sharing

Reference

---

The dataset used in indoor surface classification is presented here. If you are going to use our dataset and codes, please cite the following publication.

@article{AsiyeSurface2023,
  title={Indoor Surface Classification for Mobile Robots},
  author={Asiye Demirtas, Gokhan Erdemir, Haluk Bayram},
  journal={PeerJ Computer Science},
  volume={},
  pages={},
  year={2023}
}