Road object detection: a comparative study of deep learning-based algorithms.

This repository is an implementation of a review article on Python 3 (TensorFlow and Caffe). We train deep learning-based models (R-FCN, Mask R-CNN, SSD, RetinaNet and YOLOv4) for road object detection on BDD100K dataset.

Please find the paper here: https://link.springer.com/article/10.1007/s11042-022-12447-5.

Requirements

Name	Supported Versions
Ubuntu	18.04.5
Python	3.6
CUDA	10.1
Cudnn	7.6.4
OpenCV	4.5.0

To install requirements virtualenv and virtualenvwrapper should be available on the target machine.

Virtual Environment Creation:

# Clone repo
git clone https://github.com/bharatmahaur/ComparativeStudy.git

# Create python virtual env
mkvirtualenv ComparativeStudy

# Add library path to virtual env
add2virtualenv ComparativeStudy

Structure

├─ Mosiac                           <- Convert to mosaic images and xml files for training models
├─ R-FCN                            <- R-FCN model training and evaluation on metrices
├─ Mask R-CNN                       <- Mask R-CNN model training and evaluation on metrices
├─ SSD                              <- SSD model training and evaluation on metrices
├─ RetinaNet                        <- RetinaNet model training and evaluation on metrices
├─ YOLOv4                           <- YOLOv4 model training and evaluation on metrices
├─ LICENSE.md
└─ README.md

Dataset

Download the Berkeley Deep Drive (BDD100K) Object Detection Dataset here. The BDD dataset has the following structure:

 └── BDD100K_DATASET_ROOT
     ├── info
     |   └── 100k
     |       ├── train
     |       └── val
     |       └── test         
     ├── labels
     └── images
            └── 100K
                ├── test
                ├── train
                └── val

Mosiac Augmentation

1. Go to the mosiac augmentation folder, run mosaic_data.ipynb
2. Use the BDD100K downloaded files or try sample folder
3. Generate mosiac xml and output images like:

Training and Evaluation

To train and evaluate the model(s) use the mosiac output files or use your own custom dataset and go to the individual model folder for furthur instructions.

Trained Models

We trained these models on BDD100K, use the weights for predictions:

1. R-FCN: https://drive.google.com/file/d/11lqFSrRVDZViJCaaKFNivVTi7gPYxLlx/view
2. Mask R-CNN: https://drive.google.com/file/d/1JRm5chovHuNm4pU8czBHnEAj4SXXP2Mz/view
3. SSD: https://drive.google.com/file/d/1SCb_5z1vhTIn3pp-VeA-KmLjejbfhlfI/view
4. RetinaNet: https://drive.google.com/file/d/13P26Bb-9IiyEMx18JNlnvjrSZH-CP8x1/view
5. YOLOv4: https://drive.google.com/file/d/1k-6Y4nGnelSOO7fg6gF6R58Zbu12W4TG/view

Citation

If you use this code, please cite our paper:

@article{mahaur2022road,
 title={Road object detection: a comparative study of deep learning-based algorithms}, 
 author={Mahaur, Bharat and others},
 journal={Multimedia Tools and Applications},
 year={2022},
 publisher={Springer}
}

Contact

Please contact bharatmahaur@gmail.com for any further queries.

License

This code is released under the Apache 2.0 License.