Parking-Spot-Detection

This repository contains the code that was used to train a parking spot detection model. While the dataset is not available, you can download the trained models and test them with the server. There is also a demo mobile app (source).

Download Models

The models can be downloaded from Google Drive. The zip file contains two folders for ResNet-18 and MobileNetV3. Each folder contains a saved torchscript model and a model description file.

Test Models

To test the model, you first need to install the required dependencies:

pip install -r requirements/requirements.txt

One of the downloaded models and its description should be moved to /api:

mv /path/to/model.pt api/model.pt
mv /path/to/model_desc.json api/model_desc.json

After that you can launch the local test server:

uvicorn api.app:app

To test that server is working, you can open localhost:8000 in browser. To use the server you can go to http://localhost:8000/docs -> predict -> Try it out and choose one of the test images in /api. You can also use this command:

curl -X 'POST' \
  'http://127.0.0.1:8000/predict' \
  -H 'accept: application/json' \
  -H 'Content-Type: multipart/form-data' \
  -F 'image_file=@api/test_free_spots.jpg;type=image/jpeg'

Dataset and Model Descriptions

The models were trained on a dataset containing 5000 images (TRAIN/VAL/TEST - 4275/225/500) of roads and parking spaces from Moscow. The inference times were tested on IPhone XS Max.

MobileNetV3

python training/run_experiment.py --model_class MobileNetV3 --mobilenetv3_size=small --use_torchvision_model --data_class ParkingSpots --optimizer AdamW --lr 0.00035 --weight_decay 0.25 --gpus 1 --precision 16 --use_local --max_epochs 16 --batch_size 64 --seed 0 --num_workers 8 --color_jitter 0.3 --horizontal_flip --degrees_affine 30 --translate_affine 0.25 --scale_margin_affine 0.3 --shear_affine 5 --lr_scheduler OneCycleLR --random_erasing  --es_patience 30 --save_torchscript

	Accuracy	Loss
Train	0.7722	0.501
Val	0.6987	0.7399
Test	0.6706	0.8363

Total Avg. Time	Avg. Inference Time Only
0.482	0.06

ResNet-18

python training/run_experiment.py --model_class ResNet --resnet_type resnet18 --use_torchvision_model --data_class ParkingSpots --optimizer AdamW --lr 0.00035 --weight_decay 0.25 --gpus 1 --precision 16 --use_local --max_epochs 15 --batch_size 42 --seed 0 --num_workers 4 --color_jitter 0.3 --horizontal_flip --degrees_affine 30 --translate_affine 0.25 --scale_margin_affine 0.4 --shear_affine 5 --lr_scheduler OneCycleLR --random_erasing --use_lr_monitor --es_patience 30 --save_torchscript

	Accuracy	Loss
Train	0.7966	0.3914
Val	0.7399	0.6817
Test	0.7022	0.7719

Total Avg. Time	Avg. Inference Time Only
1.224	0.8

References

• Test Images from Wikimedia Commons:
  • test_free_spots.jpg
  • test_taken_spots.jpg
  • test_free_error.jpg