In this project, the below items are covered.

• Download dataset using the provided script(download_process.py)
• Implement a script(create_split.py) splitting the downloaded dataset into three sets(train, validation, and test)
• Implement a script(Exploratory Data Analysis.ipynb) analyzing the downloaded dataset
• Train/Evaluate the reference model using the provided scripts
• Improve the reference model by changing the training configurations
• Generate a video showing the working of the improved model

I've not modified the provieded Dockerfile. However, it seems that the ToT Dockerfile is not working properly(raising a lot of exceptions). So, after building the docker image, I've modified the below manually:

• Replace keras package: link
  • The dataset download script didn't work without this modification
• Replace function base script: link
  • The evaluation script didn't work without this modification

mkdir {PRJ_TOP_DIR}
cd {PRJ_TOP_DIR}
git clone https://github.com/aram-father/SDC-Project-1.git

cd {PRJ_TOP_DIR}/SDC-Project-1/build
docker build -t project-dev -f Dockerfile .
docker run --gpus all -v {PRJ_TOP_DIR}/SDC-Project-1/:/app/project/ --network=host -ti project-dev bash

cd {PRJ_TOP_DIR}/SDC-Project-1
mkdir ./data
python download_process.py --data_dir ./data

Exploratory Data Analysis.ipynb explores the downloaded dataset. As the below figure shows, the dataset has a lot of images containing vehicle, pedestrian, and cyclist labels.

Furthermore, I've sub-sampled the images from the dataset and visualized the label distribution as below. I could see that there is a relatively smaller number of labels for cyclists than vehicles and pedestrians.

create_split.py splits the downloaded dataset(III.3) into the below three sets.

• train: 80% of the dataset
• val(idataion): 10% of the dataset
• test: 10% of the dataset

I could see that the loss didn't reach to the plateau. So, it might be possible to improve the model by increasing the training epoch.

My finding from the result of the reference model was that the loss didn't reach to plateau. So, I've increased the trainig epoch.

I found that there is an augmentation called ssd_random_crop which is recommended for the ssd resnet. So, I've used it.

Lastly, I found that a alot of images are dark and the labels are occuluded frequently. So, I've added the below data augmentations.

• random_rgb_to_gray
• random_adjust_brightness
• random_black_patches

The resulting pipeline config file is ./experiments/experiment2/pipeline_new.config.

The below figure shows the result of the implroved model.

I've exported the improved model at ./experiments/experiment2/exported/saved_model

Here is the resulting inference video.