Skip to content

Latest commit

 

History

History
234 lines (186 loc) · 6.57 KB

README.md

File metadata and controls

234 lines (186 loc) · 6.57 KB
Raw

Revising deep learning methods in parking lot occupancy detection

Welcome to the research repo of the ''Revising deep learning methods in parking lot occupancy detection'' paper. Here we published the actual code regarding the parking lot occupancy detection problem considered in our study.

Anastasia Martynova, Mikhail Kuznetsov, Vadim Porvatov, Vladislav Tishin, Natalia Semenova.

arXiv PDF: https://arxiv.org/abs/2306.04288

Datasets

In this section, we introduce the processed versions of datasets used in our experiments: ACPDS, ACMPS, CNRPark, PKLot and SPKL.

Links to the datasets:

Dataset structure:

[DATASET_NAME]
│
└───images
│       image1.jpg
│       image2.jpg
│       ...
│
└───int_markup
│       image1.json
│       image2.json
│       ...
|
└───patch_markup
|       └───classes
|             └───Busy
|             |     image1.json
|             |     image2.json
|             |     ...
|             |
|             └───Free
|                   image3.json
|                   image4.json
|                   ...
|
│       image1.json
│       image2.json
|       image3.json
|       image4.json
│       ...
|
└───patch_splitted
|       └───train
|       |     └───Busy
|       |     |     patch1.jpg
|       |     |     patch2.jpg
|       |     |      ...
|       |     |
|       |     └───Free
|       |           patch3.jpg
|       |           patch4.jpg
|       |           ...
|       |
|       └───test
|       └───val
|
└───splitted_images
|       └───train
|       |     image1.jpg
|       |     image2.jpg
|       |     ...
|       |
|       └───test
|       └───val
|
[DATASET_NAME]_image_level_markup.json
busy_ilm_markup.json
free_ilm_markup.json
[DATASET_NAME]_dataframe.csv

For each of the datasets, we also provided visual condition labels describing the presence of the special effects in the images:

  • Sunny - sunny weather;
  • Overcast - overcast weather;
  • Rain - rainy weather;
  • Winter - winter weather, snow on the ground;
  • Fog - foggy weather;
  • Glare - glare on the image;
  • Night - night time;
  • Infrared - infrared image;
  • Occlusion (car) - cars overlap each other;
  • Occlusion (tree) - trees overlap cars and parking lots;
  • Distortion - parking lot is distorted or the image recorded by a camera with a wide angle.

Image-level annotations have the following structure:

[DATASET_NAME]_annotation_classes
│
└───CLASS_NAME_1
│       image1.jpg
│       image2.jpg
│       ...
│
└───CLASS_NAME_2
│       image3.jpg
│       image4.jpg
│       ...
│
└───CLASS_NAME_3
│       image5.jpg
│       image6.jpg
│       ...
│
...

Overall, each dataset's structure can be described as follows:

  • images folder contains the images of parking lots;
  • int_markup folder contains the intersection-level annotations of parking lots;
  • patch_markup folder contains the patch-level annotations of parking lots;
  • patch_splitted folder contains the split patches of parking lots;
  • splitted_images folder contains the split images of parking lots;
  • DATASET_NAME_image_level_markup.json file contains the image-level annotations of parking lots;
  • busy_ilm_markup.json and free_ilm_markup.json files contain the image-level annotations of free and occupied parking lots.

Custom dataset

All scripts for data preprocessing are located in the data_preprocessing folder.

To preprocess your own dataset, you need to create a folder with the name of your dataset and put the images of parking lots in the images folder. After that, you should perform the labelling procedure using the widgets from the annotators folder.

Experiments running

In this study, we explored two approaches to the parking lot occupancy detection problem: patch-based classification and intersection-based classification.

List of implemented baseline models for patch-based classification:

CNN models:

  • ResNet50;
  • MobileNet;
  • CarNet;
  • AlexNet;
  • mAlexNet;
  • VGG-16;
  • VGG-19;
  • CFEN.

Vision transformer models:

  • ViT (no prt);
  • ViT (prt);
  • DeiT (prt);
  • PiT (prt);
  • ViT (SPT LSA).

List of implemented baseline models for intersection-based classification:

  • Faster R-CNN with ResNet-50 backbone;
  • Faster R-CNN with MobileNet backbone;
  • Faster R-CNN with VGG-19 backbone;
  • RetinaNet with ResNet-50 backbone;
  • RetinaNet with MobileNet backbone;
  • RetinaNet with VGG-19 backbone.

All experiments were conducted on NVIDIA Tesla V100 GPUs.

Prerequisites

To run the code, you need to install the requirements using the following command:

pip install -r requirements.txt

Alternatively, you can create and activate the conda enviroment:

conda env create -f environment.yml
conda activate parking_research

We recommend using Python >= 3.7 and torchvision >= 1.7.0 as far as some of the architectures are not implemented in the previous versions.

Baseline models

All baseline models are located in the baselines folder. All models are implemented using PyTorch and can be trained with the help of train files by passing the model name.

Data processing

The data processing script process_data.py is located in the data_preprocessing folder. You should pass the name of your dataset as an argument:

python process_data.py --dataset DATASET_NAME

Important!: you should configure the following folder structure:

[DATASET_NAME]
│
└───[DATASET_NAME]
|
└───[DATASET_NAME]_annotation_classes

The structure of the included directories is described above in the Data section.

Contact us

If you have some questions about the code, you are welcome to open an issue or send me an email, I will respond to that as soon as possible.

License

Established code released as open-source software under the MIT license.

Citation

To be updated.

@misc{martynova2023revising,
      title={Revising deep learning methods in parking lot occupancy detection}, 
      author={Anastasia Martynova and Mikhail Kuznetsov and Vadim Porvatov and Vladislav Tishin and Andrey Kuznetsov and Natalia Semenova and Ksenia Kuznetsova},
      year={2023},
      eprint={2306.04288},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}