martin-marek / parking-space-occupancy

📄Image-Based Parking Space Occupancy Classification: Dataset and Baseline

Image-Based Parking Space Occupancy Classification

Official repository for the Image-Based Parking Space Occupancy Classification: Dataset and Baseline paper.

We introduce a new dataset for image-based parking space occupancy classification and propose a simple baseline model which achieves 98% accuracy on unseen parking lots.

In this repository, we provide:

• code to reproduce all of our results
• download links for the dataset, training logs, and model weights
• Colab notebooks to explore the dataset and models, train a model, and plot the training logs

Dataset

The dataset (called Action-Camera Parking Dataset) contains 293 images captured at a roughly 10-meter height using a GoPro Hero 6 camera. Here is a sample from the dataset:

Inference

Here's a minimal example to run inference on a trained model. For more, please see the demo notebook.

import torch, os, requests
from models.rcnn import RCNN
from utils import transforms

# create model
model = RCNN()

# load model weights
weights_path = 'weights.pt'
if not os.path.exists(weights_path):
    r = requests.get('https://storage.googleapis.com/pd-models/RCNN_128_square_gopro.pt')  
    with open(weights_path, 'wb') as f:
        f.write(r.content)
model.load_state_dict(torch.load(weights_path, map_location='cpu'))

# inference
image = torch.zeros([3, 1000, 1000])
parking_space_coordinates = torch.zeros([10, 4, 2])
image = transforms.preprocess(image)
class_logits = model(image, parking_space_coordinates)
class_scores = class_logits.softmax(1)[:, 1]

Training

To reproduce our full results from the paper, please run the train_all_models script locally. To train just a single model, please use the provided Colab notebook – Google Colab is sufficient for this.

Citation

@misc{marek2021imagebased,
      title={Image-Based Parking Space Occupancy Classification: Dataset and Baseline}, 
      author={Martin Marek},
      year={2021},
      eprint={2107.12207},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

parking-space-occupancy's Issues

How do I use my images in inference along with my rois? should it be in the same format as the dataset?

RuntimeError: CUDA out of memory

There is an issue of RAM during the training of the FasterRCNN_FPN on Colab.

How to optimise the code?