To address grid map classification problem, we construct a dataset containing 6916 (the labels including 3210 normal and 3706 abnormal )grid maps through an indoor robot vacuum cleaner. These grid maps are created with an initial size of 50m×50m. To further increase the number of training examples, we applied random rotation and offset to cropped areas of 34m×34m used as training examples. To the best of our knowledge, OGMCD is a large-scale benchmark specifically for indoor grid map classification.
- Inside of
OGMCD/python/directory runconda create -n myenv python=3.6. - Activate the virtual environment by running
source activate myenv - Install requirements from
requirements.txtby runningpip install -r requirements.txt
We train 400 epochs by Stochastic Gradient Descent (SGD)with the momentum of 0.9 and a weight decay of 1e-4. The learning rate starts from 0.01 and drops every 50 epochs. It takes about 10 hours for the network to converge on an NVIDIA GTX 2080Ti graphics card.
python train.py [OGMCD with train and val folders] train [path to weights file saves] -a [model name]For example
python train.py [OGMCD-folder with train and val folders] train ./model_save/ -a se_resnet32python test.py [OGMCD with test folders] test [path to weights file] -a [model name]For example
python test.py [OGMCD with test folders] test se_resnet32.pth -a se_resnet32You may download the dataset reported in the paper from Google Drive or the Baidu Netdisk
| Google Drive | Link |
|---|---|
| Baidu Netdisk | Link |
Baidu Netdisk eval code：yyvs