DT_objdet_trian

About the Directory

This directory is used for generating and formatting a Duckietown-specific dataset which is then used to train a YOLOv5 model. Once trained the YOLOv5 model can be run using the DT_objdet repository on the same Github account. The dataset comprises real and simulated images from the publically available Duckietown dataset. Both real and simulated images are collected within the duckietown-lx environment. The model is trained using the Google Colab script found in this repository

NOTE:

If you would like to skip the training, use the following DT Token: dt1-3nT8KSoxVh4MnDRxovGLkXZDhPpgc4SzasJBTSxbRUfDguS-43dzqWFnWd8KBa1yev1g3UKnzVxZkkTbfeFCAD1kMCPQvvSVDYPfoXapvF29wVgdC7 in the constants.py file from the DT_objdet.

Prerequisites

The list below states the prerequisites to use this directory.

1. Cloned duckietown-lx repository.

Instructions

1. Replace the data_collection.py and setup_activity.py files in the duckietown-lx/object-detection/ directory with the two files with the same names found in the replace folder in this directory. These files contain modifications that extend the dataset to include labels for duckiebots. The main difference between the data_collection.py files is that the modified version in this directory file saves segmented images to the assets directory. The setup_activity.py file has modifications that are necessary to add duckiebot labels.

2. Follow the instructions in the duckietown-lx/object-detection-lx\setup.ipynb notebook to download the real dataset. Ensure that your terminal is cd'ed into the duckietown-lx/ and then run the data_collection.py using the following two lines below:

   dts code build
   dts code workbench --simulation --launcher data-collection

   The simulation will automatically close when the specified number observations have been obtained.

3. Copy the images in the asset directory of the duckietown-lx\object-detection directory across to this directory.

4. cd into this directory and run the sim_image.py and real_image.py files. Note that you will have to manually add the path to your directory in the DATASET_DIR variable for each of these files. This should move all images and labels directory to the train and validation folders.

5. Run the dataset_zip.py. This should zip up the duckietown_object_detection_dataset. Copy the dataset to a Google Drive

   Please be aware that

   • Do not rename the dataset zip file
   • The file should be uploaded to the out-most "My Drive" area

6. Use the DT_training.ipynb notebook in this directory to train the YOLOv5 model. The notebook walks you through the procedure, after training there should be a folder in your Google drive where you can look at the training results.