Downloads · 30 days
30
31% of all-time downloads
Paulyang5049/tennis-ball-yolo26s-experimental-v0.2.0
tennis-ball-yolo26s-experimental-v0.2.0 is a object detection model from Paulyang5049. Use it when you need objects located in an image. It is set up for ultralytics. The card lists the license as agpl-3.0.
This repository contains the experimental tennis-ball checkpoint released with Tennis AI Local v0.2.0. It is published as a reproducibility baseline and failure-analysis artifact.
Downloads · 30 days
30
31% of all-time downloads
All-time downloads
98
Public
Repo size
80.4 MB
Likes
0
Public
Click a slice to open those files.
.pt80.4 MB · 100%
From the Hugging Face model README
This repository contains the experimental tennis-ball checkpoint released with Tennis AI Local v0.2.0. It is published as a reproducibility baseline and failure-analysis artifact.
This candidate failed its untouched-test promotion gate and is not the default model in Tennis AI Local. At confidence 0.1 and IoU 0.5, it produced 0 true positives, 34 false positives and 50 false negatives on the held-out test split. The COCO sports-ball baseline produced 2 true positives, 21 false positives and 48 false negatives on the same split.
Do not use this checkpoint for line calls, officiating, safety decisions, or claims of validated tennis-ball accuracy.
tennis ball7fc1e1cd507381e7b100043012d9db7805a59726f16cf1827a6c436b7fd22c97Ultralytics' best-checkpoint validation reported precision 0.187, recall 0.200,
mAP50 0.0616 and mAP50-95 0.0123. The fixed comparison stored in
manifest.json used confidence 0.1 and IoU 0.5.
The checkpoint was fine-tuned on version 6 of the Tennis Ball Detection dataset by Viren Dhanwani, licensed CC BY 4.0. The audited split contained 468 training, 60 validation and 50 test images. No exact decoded-pixel duplicates were found. Source-video identities and possible near-duplicates remain unverified, so match-level independence is not established.
The dataset itself is not included here. Attribution is retained in
DATASET_ATTRIBUTION.txt.
Install the tested Ultralytics version and Hugging Face Hub client:
pip install "ultralytics==8.4.140" huggingface_hub
from huggingface_hub import hf_hub_download
from ultralytics import YOLO
weights = hf_hub_download(
repo_id="Paulyang5049/tennis-ball-yolo26s-experimental-v0.2.0",
filename="weights.pt",
)
model = YOLO(weights)
results = model.predict("frame.jpg", imgsz=1280, conf=0.1)
Ultralytics .pt checkpoints use PyTorch serialization. Download only from a
repository and revision you trust, and verify the checksum before loading.
The test result shows that this checkpoint does not generalize reliably to the held-out images. Small and far-court balls, motion blur, occlusion, compression, lighting changes and non-ball bright objects remain difficult. The split uses filename-family grouping because original source-video identities were not available. No validated phone-video, doubles, broadcast-sequence or physical- device accuracy result is claimed.
The fine-tuned YOLO26s checkpoint is released under AGPL-3.0. The source code, training notebook, evaluation method and project licence are available in the Tennis-AI-Tracker repository. The training dataset retains its CC BY 4.0 terms and attribution. Commercial users should review the current Ultralytics licensing terms for their intended use.