imaging_training_2024

Synthetic data

Generation

Currently, the code only works on Vince's computer. The many attempts to run it elsewhere have thus failed. It's available at https://github.com/uci-uav-forge/UAV-Forge-Imaging-Sim for reference.

Retrieval

The generated data can be downloaded from https://drive.google.com/drive/folders/1bFDhUh6WVlGPChKcFt_PdUnMzQb1qoSw. The relavent files are train.tar.gz, semantic_segmentation_labels_.tar.gz, semantic_masks.tar.gz, bbox_location_labels.tar.gz, and bbox_location_labels.tar.gz. Note that the last two are nested in the bounding_box_labels folder.

Usage

Interface Setup

The interface written in sim_data_interface assumes that the folders are flattened, such that the structure is:

data [self-defined]
├── bbox_class_labels
├── bbox_labels
├── semantic
├── semantic_segmentation_labels
├── train

The default names can be changed in sim_data_interface/config.py and can also be overridden in the Dataset constructor. Note that the above names are not the same as the archive names; they're the names on the directories inside the archive.

Interface Usage

The interface is written in sim_data_interface/dataset.py. The Dataset class is the main interface. There is an example at the bottom of the file.

Latest dataset link (january): https://drive.google.com/drive/folders/1bFDhUh6WVlGPChKcFt_PdUnMzQb1qoSw

Troubleshooting

_imagingft Import Error

• Install libfreetype: sudo apt-get install libfreetype6-dev
• Create a a new environment
• Reinstall the packages with pip install -r requirements.txt --no-binary :all
  • --no-binary :all forces recompilation, which can help.

YOLO: Dataset Not Found

• Check the path of the subfolders in data.yaml.
  • If path is defined, it is relative to the CWD, not the file itself. Use absolute paths to be sure.
  • The other path parameters are relative to the path parameter, if defined.

YOLO: Main thread not in main loop

• This is a TKinter error caused by threaded invocation of matplotlib visualization.
• Solution: Add matplotlib.use('Agg') in ultralytics/yolo/utils/plotting.py to disable the visualization.
  • This is janky, but it works.