A dedicated YOLOv11s detector for the shuttlecock — a small, fast, low-contrast object that general detectors miss. Trained and run at high resolution (imgsz=1280) to recover per-frame shuttle positions for...
Source du modèle
Extrait de la source
A dedicated YOLOv11s detector for the shuttlecock — a small, fast, low-contrast object that general detectors miss. Trained and run at high resolution (imgsz=1280) to recover per-frame shuttle positions for...
Sources
1 sourceVérifié 1 août
Artefacts du modèle
1 artefactExtraits de sources
2 extraits--- license: agpl-3.0 library_name: ultralytics pipeline_tag: object-detection tags: - ultralytics - yolo11 - yolov11 - object-detection - small-object-detection - badminton - sports-analytics --- # Badminton Shuttlecock Detector (YOLOv11s, fine-tuned) A dedicated YOLOv11s detector for the **shuttlecock** — a small, fast, low-contrast object that general detectors miss. Trained and run at **high resolution (`imgsz=1280`)** to recover per-frame shuttle positions for trajectory building and shot detection. - **Task:** object detection (`task=detect`) - **Classes:** `Shuttlecock` (1 class) - **Base checkpoint:** `yolo11s.pt` (Ultralytics) - **Framework:** Ultralytics 8.4.87 ## Intended use Per-frame shuttle localisation on single-camera broadcast footage. Output positions form the shuttle trajectory that a downstream angle-change detector uses to trigger shot events. **Run at `imgsz=1280`** — the model was trained at 1280 and small-object recall drops sharply at lower resolutions. ## How to use ```python from ultralytics import YOLO from huggingface_hub import hf_hub_download w = hf_hub_download("<your-username>/badminton-shuttlecock-yolov11", "shuttle_yolo11s.pt") model = YOLO(w) res = model.predict("frame.jpg", imgsz=1280, conf=0.25)[0] if len(res.boxes): cx, cy = res.boxes.xywh[0][:2].tolist() # shuttle centre (pixels) ``` ## Training | | | |---|---| | Base model | `yolo11s.pt` | | Epochs | 100 | | Image size | 1280 | | Batch | 8 | | Dataset | Roboflow Universe — *Shuttlecock* (`mathieu-cartron`) | | Dataset link | https://universe.roboflow.com/mathieu-cartron/shuttlecock-cqzy3 | ## Evaluation (validation split, from the training checkpoint) | Metric | Value | |---|---| | Precision (box) | 0.734 | | Recall (box) | 0.631 | | mAP@50 (box) | 0.709 | | mAP@50-95 (box) | 0.294 | > Small-object detection is inherently hard; the modest mAP is expected. In practice, running > at `imgsz=1280` on broadcast frames yields dense-enough per-frame coverage for trajectory > reconstruction (e.g. ~211/250 frames on a sampled rally segment in our pipeline). Coverage > varies with footage quality and is not a validation metric. ## Limitations - Single class; broadcast viewpoint only. - Recall degrades at lower inference resolutions and on motion-blurred / occluded frames. - Small validation split (see source dataset); treat metrics as in-domain. ## License...
Source context: 26 downloads · 0 likes · Pipeline object-detection · Library ultralytics · Repo Huydinh1205/shuttle_yolo11s