Boggart

Visualization

visualization.mp4

Environment/Repository Setup Instructions

Python Environment

cd boggart
sudo apt install python3.7
python3.7 -m venv env
source env/bin/activate
pip install --upgrade pip
pip install opencv-python tqdm tensorflow munkres 
pip install mongoengine

In configs.py, set BOGGART_REPO_PATH to the location of the Boggart repository.

Hwang Frame Extraction Repository

git clone https://github.com/neilsagarwal/hwang
sudo apt-get install git cmake libgoogle-glog-dev libgflags-dev yasm libx264-dev build-essential wget unzip autoconf libtool
cd hwang
bash deps.sh
export LD_LIBRARY_PATH=<PATH_TO_BOGGART_REPO>/hwang/thirdparty/install/lib:$LD_LIBRARY_PATH
mkdir build
cd build
cmake ..
make -j64
cd ..
bash build.sh
export LD_LIBRARY_PATH=<PATH_TO_BOGGART_REPO>/hwang/python/python/hwang/lib:$LD_LIBRARY_PATH
 
cd ..

For mAP Evaluation

git clone --depth 1 https://github.com/tensorflow/models
cd models/research
protoc object_detection/protos/*.proto --python_out=.
cp object_detection/packages/tf2/setup.py .
python -m pip install .
cd ..

Data Setup Instructions

Video Data Setup

Videos (and data generated during execution) will be stored at <BOGGART_REPO_PATH>/data/. Make sure to set this path in configs.py (see the main_dir property).

Videos are expected to be split up by hour and then stored in ten-minute chunks. For example, the first ten minute chunk of hour 10 of the auburn_first_angle video dataset would be located at <BOGGART_REPO_PATH>/data/auburn_first_angle/video/auburn_first_angle10_0.mp4.

Example file structure for data/:

• boggart/
  • data/
    • auburn_first_angle10/
      • video/
        • auburn_first_angle10_0.mp4
        • auburn_first_angle10_1.mp4
        • auburn_first_angle10_2.mp4
        • auburn_first_angle10_3.mp4
        • auburn_first_angle10_4.mp4
        • auburn_first_angle10_5.mp4
    • auburn_first_angle11/
      • ...

Model Inference Data Setup

Boggart's current implementation requires that model results are already generated and saved into MongoDB. The repository contains a helper script to load model inference results into MongoDB. load_inference_results_into_mongodb.py requires that inference results are stored in per-hour chunks. For example, the inference results for running YOLOv3 (trained on the COCO dataset) for hour 10 of the auburn_first_angle video dataset should be located at <BOGGART_REPO_PATH>/inference_results/yolo3-coco/auburn_first_angle/auburn_first_angle10.csv .

To set up MongoDB, run:

sudo apt install -y mongodb

Add 'directoryperdb=True' to /etc/mongodb.conf.

sudo mongod --config /etc/mongodb.conf

Then, update ml_model, video_name and hour in load_detections_into_mongodb.py. Running this script will then load that hour's worth of inference results into the database.

Run Boggart

Instructions to execute Boggart's ahead-of-time and query-time processing can be found in run.py.