AIDOVECL:

AI-generated Dataset of Outpainted Vehicles for Eye-level Classification and Localization

Installation Guide

To set up the environment for AIDOVECL, follow these steps:

First, ensure you have conda or miniconda installed. Then, create a new conda environment using the provided aidovecl-env.yml file.

conda env create -f aidovecl-env.yml
conda activate aidovecl

Run the following command to add the environment to Jupyter:

python -m ipykernel install --user --name=aidovecl --display-name="Python (aidovecl)"

Reproducing Figures

To produce the figures of the paper, after activating aidovecl environment, execute the jupyter notebooks.

Note 1: Please be patient during the initial run as the detection and inpainting models are being downloaded.

Note 2: Despite setting random seeds, 100% reproducibility is not guaranteed. For more information, refer to cuBLAS and PyTorch.

An Overview of the Package:

• detect.py: Detects vehicles and creates seed images.
• outpaint.py: Outpaints the seed images.
• backdrop.py: Generates background images.
• aidovecl.py: Generates AIDOVECL, backgrounds, and augments real data with them.
• yolo.py: Trains and tests YOLO on datasets
• utils.py: Provides utilities for the above files.

For use cases, refer to jupyter notebooks.

Downloading Datasets for Vehicle Classification and Localization

Download the dataset from the following source: AIDOVECL Dataset.

After downloading, extract the zipped datasets to the datasets folder of the repository. The structure should look like this:

datasets/
    ├── real
        ├── bus
        ├── coupe
        ├── minibus
        ├── minivan
        ├── pickup
        ├── sedan
        ├── suv
        ├── truck                                   
        └── van
    ├── real_seeded_split
        ├── images
        ├── labels
        └── real_seeded_split.yaml
    └── augmented
        ├── images
        ├── labels
        └── augmented.yaml

The jupyter notebook demo-fig-4.ipynb demonstrates how to use real dataset to generate real_seeded_split and augmented datasets. It also showcases training and testing YOLO on them. Note that real dataset consists of collected images with no annotations, whilereal_seeded_split includes selected images for outpainting that are also annotated and split into train, val, and test folders.