TrackNetV2: Efficient TrackNet

⚙️ 1. Install

System Environment

• Windows 10
• NVIDIA RTX3060
• Python3.5.2 / git / PYQT5 / OpenCV / pandas / numpy / PyMySQL
• TensorFlow 1.13.1/keras 2.2.4/Opencv 4.1.0/CUDA 10.1

Package

• First, you have to install cuda, cudnn and tensorflow, tutorial: https://medium.com/@zhanwenchen/install-cuda-and-cudnn-for-tensorflow-gpu-on-ubuntu-79306e4ac04e

    $ sudo apt-get install git
    $ sudo apt-get install python3-pip
    $ pip3 install pyqt5
    $ pip3 install pandas
    $ pip3 install PyMySQL
    $ pip3 install opencv-python
    $ pip3 install imutils
    $ pip3 install Pillow
    $ pip3 install piexif
    $ pip3 install -U scikit-learn
    $ pip3 install keras
    $ git clone https://nol.cs.nctu.edu.tw:234/open-source/TrackNetv2
  

🎬 2. Prediction for a single video

Generate the predicted video and the predicted labeling csv file

You can predict coordinate of shuttlecock for a single video with:

python3 predict.py --video_name=<videoPath> --load_weights=<weightPath>

Just put the video path you want to predict on option <videoPath> . We provide the pretrain model weights model_33 for 3_in_1_out version, and model906_30 for 3_in_3_out version.

Show the predict trajectory

After predict.py, you will have the predicted labeling csv file. You can apply show_trajectory.py to generate ball's trajectory video for fancy purpose.

python3 show_trajectory.py <input_video_path> <input_csv_path>

For example:

• <input_video_path> = 1_01_00.mp4
• <input_csv_path> = 1_01_00_predict.csv

python3 show_trajectory.py 1_01_00.mp4 1_01_00_predict.csv

🛠️ 3. How to train your own data

Step 1 : Generate the frame of video

python3 Frame_Generator.py <videoPath> <outputFolder>

<videoPath> is the video path you want to train, and <outputFolder> is the output directory you want to store the frames under.

For example:

• <videoPath> = 1_01_00.mp4
• <outputFolder> = frame

python3 Frame_Generator.py 1_01_00.mp4 frame

Step 2 : Preprocess the labeling csv file

We will need .csv file generated by [our labeling tool].

Step 3 : Generate training data

python3 gen_data.py --batch=<batchSize> --label=<csvFile> --frameDir=<frameDirectory> --dataDir=<npyDataDirectory>

<csvFile> is the .csv file after apply Rearrange_Label.py at second step, <frameDirectory> is the frame folder of the video at first step, and <npyDataDirectory> is the output directory you want to store the training data under.

For example:

• <batchSize> = 1000
• <csvFile> = 1_01_00_ball.csv
• <frameDirectory> = frame
• <npyDataDirectory> = npy

python3 gen_data.py --batch=1000 --label=1_01_00_ball.csv --frameDir=frame --dataDir=npy

Step 4 : Start training TrackNetV2

python3 train_TrackNet.py --save_weights=<weightPath> --dataDir=<npyDataDirectory> --epochs=<trainingEpochs> --tol=<toleranceValue>

<weightPath> is TrackNetV2 weight after this training, <npyDataDirectory> is the directory of the .npy training data at third step, and <toleranceValue> means tolerance value of true positive.

For example:

• <weightPath> = mymodel
• <npyDataDirectory> = npy
• <trainingEpochs> = 30
• <toleranceValue> = 4

python3 train_TrackNet.py --save_weights=mymodel --dataDir=npy --epochs=30 --tol=4

Step 5 : Retrain TrackNetV2 model

If you want to retrain the model, please add load_weights argument.

python3 train_TrackNet.py --load_weights=<previousWeightPath> --save_weights=<newWeightPath> --dataDir=<npyDataDirectory> --epochs=<trainingEpochs> --tol=<toleranceValue>

<previousWeightPath> is the model weights you had trained before.

For example:

• <previousWeightPath> = mymodel
• <newWeightPath> = mynewmodel
• <npyDataDirectory> = npy
• <trainingEpochs> = 10
• <toleranceValue> = 4

python3 train_TrackNet.py --load_weights=mymodel --save_weights=mynewmodel --dataDir=npy --epochs=10 --tol=4

📔 5. Provide the performance information

python3 accuracy.py --load_weights=<weightPath> --dataDir=<npyDataDirectory> --tol=<toleranceValue>

accuracy.py provide following version:

• Number of true positive
• Number of true negative
• Number of false positive
• Number of false negative
• Accuracy
• Precision
• Recall

<toleranceValue> means tolerance value of true positive. <npyDataDirectory> is the directory of the .npy data you want to do accuracy measure.