broughtong/car-detector

Self-learning Car Detector

Arranging the data

Place the rosbags into data/rosbags Place the gt into data/gt

Set-up your system

You will need to have tf_bag package installed and sourced, from here: https://github.com/IFL-CAMP/tf_bag For the rest of the dependencies, there are conda environments provided in the conda folder. The ros environment is used for everything, except for running lanoising You can create the environments like this: mamba env create --file lanoise.yml

Extract the rosbags

Run preprocessing/rosbag_extractor.py to extract the rosbags Run preprocessing/tf_extractor.py to extract the transforms

Perform simple car detection

Run supervisor/detector.py to run the method

Prepare Annotations

Run supervisor/temporal.py to run the temporal filter

You will need the weather simulator network models, explained on the simulator page. The origal page is here https://github.com/cavayangtao/lanoising And a mirror is provided here: https://github.com/broughtong/lanoising Copy the models into supervisor/lanoising/models/

Switch to the lanoise conda environment. For me, I need to make sure PYTHONPATH is blank, activate the conda environment, and then source ros, in that order. Run the lanoising weather simulator in supervisor/lanoising/lanoising.py

Run supervisor/annotator.py to create annotations. Run supervisor/datasetDivision.py to break the images into training/testing/evaluation subset by rosbag (all bags with gt go to eval)

Run networks/generate_data.py Train pointnet using python train_segmentation.py and python train_segmentation.py --lanoise Optionally add the feature transform flag --feature_transform Train unet using python trian_unet.py Optional flags --num_channels 3 (1/2/3) --lanoise and --numc 2 (2/3)

Train the networks

The MaskRCNN network can be trained by running networks/maskrcnn/training.py Alternatively, you can run the files in network/ to train the other networks.

Evaluation

For pointnet use networkEval.py Make sure the feature transform flag is set correctly. For each network, go into the folder and run the associated evaluation script. To generate the graphs, run the scripts in the evaluation folder.