Code & dataset repository for the paper: BlurBall: Ball detection with blur estimation
This repo is forked from WASB: Widely Applicable Strong Baseline for Sports Ball Detection and Tracking We added the training scripts and and other modifications.
The table tennis ball dataset includes both the positions of the balls (either midpoint or endpoint) and the motion blur associated (length and orientation).
It can be dowloaded from here: NextCloud
All trained model weights for BlurBall, WASB, TrackNetv2, ResTrackNetv2, BallSeg, DeepBall, DeepBall large and Monotrack are available here: Nextcloud
Because the BlurBall is multiple input multiple output, it is quite sensitive to duplicated frames. This often happens on online recordings where videos recorded at 25fps are encoded at 30fps. It will generate a directory with the unique frames at the same location as the input video.
Run inference on a video:
python main.py --config-name=inference_<model> detector.model_path=<path to corresponding model> +input_vid=<path to vid>
To come