Flowlets

Converts KITTI tracklets into vectors denoting flow, for object tracking

by Alvin Wan

Flowlets are vectors between frames, from the center of a bounding box at time t to a bounding box at time t+1. Each flowlet is associated with an object, for all frames the involved object is detected.

The repository is written in Python 3.

Installation

First, clone the repository. (Optional) Create a virtual environment, and launch the environment.

virtualenv flowlet --python=python3
source flowlet/bin/activate

Let $FHOME$ denote the repository root. Install all Python dependencies, from the repository root.

cd $FHOME
pip install -r requirements.txt

If you're only looking to convert tracklets to flowlets, run the following download script. Otherwise, please see the KITTI website for a full dataset download.

bash tracklet_calib_data_downloader.sh

Finally, use python flowlets.py per the usage instructions below. You may additionally use python flowlets.py --help to retrieve instructions.

Usage

The script is contained in the repository root. To, for example, convert all tracklets into flowlets, run the following, where /KITTI_raw is the directory containing our raw KITTI data, and ~/flowlets should contain all outputs.

python flowlets.py KITTI /KITII_raw --out=~/flowlets

Usage:
    flowlets.py <path> [options]
    flowlets.py KITTI <kitti_dir> [options]
    flowlets.py KITTI <kitti_dir> <date> <drive_id> [options]

Options:
    --d=<dims>          Number of dimensions [default: 2]
    --out=<dir>         Directory containing outputted files. [default: ./out]
    --mode=(obj|frame)  Output one file per object or one file per frame. [default: frame]
    --columns=<columns> Specify order of columns