Tianjie Zhang(tjzhang@u.boisestate.edu),
Amanda Jo Mullins,
Steven Kim,
Yang Lu
We mainly used Generative adversarial network(GAN) to augment images in this competition. In the whole competing process, we try to perform some image-preprocess such as Histogram equalization, and some filters including Median, Gamma, Gaussian etc. However, all the image preprocess methods showed an accruacy decrease in the final result. Also, we tried to use the traditional image augmentation methods like crop, noise, flip. These methods did not improve the final result. We were successful on utilizing Generative adversarial network to produce fake road images. It produced a variety of different images depending on our training data. Thankfully, we got good results which proves that it is a powerful method to augment the preexisting pavement images. Our final accuracy is around 0.633.
- Create the environment with conda env create:
conda env create -f environment.yml You can also update an environment to make sure it meets the current requirements in a file:
conda env update -f environment.yml- Annotate the second batch of training Data released from DSPS using CVAT. Put the images and annotation file into the td4 folder under cvat folder.
- This is the my team's annotation file for the second batch of training data.
- Run the GAN.ipynb (Generative adversarial network) code to produce some fake road images.
- To reproduce what we have done, download all the necessary libraries (make sure you have done the step.1), and change the value of dataset = "" to your desirable location.
- Training is performed through the images in the ng folder, and validating is performed through the images in the ok folder under the dataset directory.
- We suggest to change the parameters n_epochs, n_cpu, img_size, sample_interval according to your system's specification and the time you have, but we highly suggest you to keep the rest of the parameters.
- We choose the best 100 fake images to train, then annotate these images also using CVAT. Put the images and annotation file into the td5 folder under cvat folder.
- This is the link for the fake road images we choose from the all produced images.
- This is the annotation file for the 100 fake images.
- Run the dsps_main.ipynb (the Yolo model provided by the DSPS)