radcog

radball ai recognition of the objects in a radball field/game

Requirements

Label Studio

used for labeling data, not strictly needed though

Other Packages

These Packagenames are for arch-linux(some AUR), but should be similar for others

• python-gitpython
• python-matplotlib
• python-numpy
• python-opencv
• python-pillow
• python-psutil
• python-yaml
• python-requests
• python-scipy
• python-thop
• python-torchvision-cuda
• python-pytorch-cuda
• python-tqdm
• python-ultralytics
• python-lap
• python-opencv

Usage

Label Data

• install label studio
• import videos (store those video files somewhere, where you can access them later)
• mark them (dont forget to check if you disabled automarking, when you only labeled part of a video)

convert Label Studio to YOLO

• Export in JSON-MIN format
• OPTIONAL: If you only want to export a single video (or a few) use the API:

curl -X GET "http://localhost:8080/api/projects/{project_id}/export?ids[]={task_id}&exportType=JSON_MIN" \
     -H 'Authorization: Token YOUR_API_TOKEN' \
     --output 'annotated_video.json'

• git clone https://github.com/hidenorly/video_annotations_to_yolo.git
• python ls2yolo.py --json_path JSON_PATH --video_path ~/.local/share/label-studio/media/upload/PROJECT_ID -o DATA_SET_PATH
• OPTIONAL: play with the parameters and number of augmentations in augment.py, change the input and output path, then run it to synthetically create more pictures (with augmentations)
• set images_dir, labels_dir and output_dir in split_data.py and run it (output_dir should be inside datasets/

Train the model

• modify data.yaml to the correct path (datasets/DATA_SET_PATH)
• modify train.py input model and if needed other parameters, then run it

Use the model

• use main.py -> detect all images in a directory (input/)
• use main2.py -> detect a video (adjust video_path)

Models

Model	Base	imgsz	box_loss	cls_loss	dfl_loss	Box(P	R	mAP50	mAP50-95	Images	Description
radcog-0.5.4	yolo11s	1028	?	?	?	?	?	?	?	454	more annotation data than 0.5.1, no augmentations
radcog-0.5.5	yolo11s	1028	0.7568	0.5142	0.8323	0.994	0.926	0.981	0.826	2025	more annotations than 0.5.4, no augmentations

Model	Base	imgsz	box_loss	cls_loss	dfl_loss	Box(P)	R	mAP50	mAP50-95	Images
First try with bad annotations (red balls)
radcog-0.1.4	yolo8l	512	?	?	?	?	?	~0.5	~0.2	~5000
Second try with 2600 high quality black balls and images from radcog-0.1
radcog-0.2.1	yolo11s	512	1.583	1.334	0.9287	?	?	0.617	0.345	~8000
Upping quality with higher res and other small adjustments
radcog-0.2.2	yolo11s	1028	1.449	1.576	1.245	?	?	0.657	0.414	~8000
More epoch waiting, as 0.2.2 doesn’t seem finished with patience=5
radcog-0.2.3	radcog-0.2.2	1028	1.415	1.346	1.174	?	?	0.68	0.426	~8000
Only use perfect black ball annotations
radcog-0.3.1	yolo11s	1028	?	?	?	?	?	?	?	2620
Use 1.0 augmentations (x8) on the 2620 pictures and spin down the low learning rate etc. from 0.3. Not finished training but stopped after realizing the training data was bad. Does not recognize black balls at all, despite "good" values.
radcog-0.4.2	yolo11s	1028	1.506	1.307	1.054	0.851	0.697	0.825	0.511	22900
Realized all conversions were completely wrong, making most of the annotated pictures bad. Fixed the conversion and reannotated around ~100 pictures for a first try (no augmentations). Instantly better than every model so far.
radcog-0.5.1	yolo11s	1028	1.882	1.212	0.9744	1	0.787	0.892	0.448	120
Use radcog-0.5.1 data with more conservative augmentations (x8) than in 0.4.2.
radcog-0.5.3	yolo11s	1028	1.544	0.8294	1.026	0.918	0.777	0.856	0.462	~1000