Model-Assisted Parking Lot Labeling

This repository contains code that was used to train a PointNet++ bounding box regression model on data that was collected using the parking lot protocol of Paved2Paradise. The code does not run as is. You will need to collect your own parking lot data and modify the prepare_datasets.py and upload_segments_labels.py scripts to reflect the idiosyncrasies of your datasets. The model can be trained on a small number of labeled sequences (e.g., one), and then used to predict bounding boxes for the remaining samples, which can then be refined in Segments.ai.

Installing the necessary Python packages

pip3 install -r requirements.txt

Generating the training/test data

python3 prepare_datasets.py

Training the PointNet++ model

Training the model requires PyTorch3D, which is easiest used in a container, so start the container:

# The path to this repo.
REPO_PATH=/home/michael/Projects/parking-lot-pointnetplusplus
# Should be the same as the DATASETS_PATH variable in env_vars.py.
DATASETS_PATH=/media/michael/Extra/aa_field_day
docker run -v ${REPO_PATH}:/workspace/pointnetplusplus -v ${DATASETS_PATH}:/workspace/datasets --runtime nvidia -it pytorch3d

After you start the container, log in to Weights & Biases:

wandb login --host <your_wandb_url>

and then train the model:

cd pointnetplusplus
python3 train.py

Generating the bounding box predictions

python3 predict_bboxes.py <run_id>

where <run_id> is the ID associated with the Weights & Biases run, e.g.:

python3 predict_bboxes.py 2gyfwhhj

To upload the bounding box predictions to Segments.ai, run:

python3 upload_segments_labels.py [visualize]

where visualize is an optional argument that when included will visualize the bounding box predictions.