DuckieRoadNet

This is the spcn_hackthon_2020 solvation of AYSAT team for predicting duckietown road types.

Road types explained here: https://docs.duckietown.org/DT19/opmanual_duckietown/out.pdf#page=4

Product Value

Done

• Fast and simple road classification model
• Simple markup tool (python script)
• Pretrained models for testing
• Model predictions are better, than random dice roll
• Good model prediction accuracy ~ 82-84%

To do

• Unbalanced dataset -> some classes solution detects with high acc, some with bad
• Unclear rule of classification, what is the next tile?
• Small Dataset for classification

Demonstration

Algorithm working on video from duckietown server: https://youtu.be/Ve79IPrYC6c

Model makes base prediction -- next tile class (on the left of "/"). Also, we print second-prioritized predicted class, cause our dataset not balanced && trained mostly on "straight line" tiles (on the right of "/").

Run

Craete python venv, install requirements && download demo videos:

pip3 install --upgrade pip
pip3 install -r requirements.txt
python3 generate_dataset.py 

For launch prediction algorithm on your video:

python3 predict_roadtype.py -i ./video.mp4 -o ./out.mp4 --canny --cpu
-i -- path input
-o -- path output video
--canny -- outout video processed with canny
--cpu -- take cpu-trained model

Create your own dataset

If you want landmark video for your own dataset (we hope you want, because we don't), use this script (doesn't needed for demo) Launch landmarking script:

cd src
python3 markup_dataset.py -p ./data/video1/ 
// -p <relative path to data directory, that contains imgs>
// -o <JSON markup filename>

HackathonTeam

• Gosudarkin Yaroslav
• Gavrilov Andrey
• Glazunov Sergey
• Tokarev Andrey
• Gizzatov Amir