Stage II - Players Detection & Action Recognition

University of Central Asia⛰️

🏃‍♂️ How to Run

1. Clone this repository

git clone https://github.com/shukkkur/VolleyVision.git

2. Install the requirements

cd VolleyVision\Stage II - Players & Actions
pip install ultralytics

3. You can run using Python script main.py

1. python main.py --model actions/yV8_medium/weights/best.pt --input_path assets/actions.jpg --output_path Output/actions.jpg --show_labels
2. python main.py --model players/yV8_medium/weights/best.pt --input_path assets/players.jpg --output_path Output/players.jpg

4. Since ultralytics provides simple CLI you can also use it

1. yolo predict model=actions\yV8_medium\weights\best.pt source=assets\actions.jpg show_conf=False show_labels=True
2. yolo predict model=players\yV8_medium\weights\best.pt source=assets\players.jpg show_conf=False show_labels=False



i.	ii.

Action Recognition + Temporal Filtering = Event Detection

For event detection I used the action recognition model, but now instead of drawing the bounding boxes on every frame, we intoduce temporal information using a list (deque). So, basically we store the last N frames' predictions and declare that an event occured only if a certain action has been predicted several times. Gosh, my explanation is terrible, please refer to below examples.

# `-` means no detections
# `()` braces are the sliding window
# let's choose sliding_window of size `5`
# and threshold number of `3`, meaning we need 3 detections for an event to be declared

predictions = [-, -, spike, -, spike, spike, -, -, spike, spike]

[(-, -, spike, -, spike), spike, -, -, spike, spike]  # only two detections, no event
[-, (-, spike, -, spike, spike), -, -, spike, spike]  # yeah, we've got 3 detection within the window, draw "SPIKE" on top
[-, -, (spike, -, spike, spike, -), -, spike, spike]  # keep drawing/printing/declraing "SPIKE"
...

I really hope this make sense now. Anyway, this simple, yet amazing idea belongs to my former mentor at BallerTV - Paul Kefer. Thank you for your patience))).

1. To run the event detection, use sliding_window.py script.

!python sliding_wndow.py --model actions/yV8_medium/weights/best.pt --input_path "assets/rallies/rally.mp4" --output_path Output/event_detection.mp4 --conf 0.5

2. The effectiveness of this approach depends on confidence level (--conf), sliding window size and threshhold, so for experimental purposes use sliding_window_verbose.py. In addition to event, it also draws [frame_number] (who's action/yellow circle) class_name (confidence)
3. --gpu - for faster inference. Can be added to other scripts (do yourself). And --imgsz - to prevent YOLO resizing the input to 640x640 but instead to larger/smaller custom size.

!python sliding_wndow_verbose.py --model actions/yV8_medium/weights/best.pt --input_path "assets/rallies/rally.mp4" --output_path Output/event_detection.mp4 --conf 0.4 --gpu --imgsz 1920 1080

sliding_window.py	sliding_window_verbose.py

For any additional quesitons feel free to take part in discussions, open an issue or contact me.