Chess Piece Detectors (YOLOv8 — pico / n / s)
Three YOLOv8 detectors trained on cell-box annotations derived from the synthetic
koryakinp/chess-positions
dataset (12 piece classes). Companion to the custom CNN classifier at
honi05/chess-piece-cnn.
| Variant | Origin | Params | mAP@50-95 | Occupied-cell acc | Latency ms/board | Disk MB |
|---|---|---|---|---|---|---|
yolo_pico/best.pt | scaled YAML, from scratch | 184,300 | 0.255 | 0.2073 | 5.49 | 0.57 |
yolov8n/best.pt | pretrained | 2,692,548 | 0.995 | 1.0000 | 4.35 | 5.59 |
yolov8s/best.pt | pretrained | 9,843,604 | 0.995 | 1.0000 | 6.07 | 19.92 |
Classes (0-11): wP, wN, wB, wR, wQ, wK, bP, bN, bB, bR, bQ, bK
Usage
from huggingface_hub import hf_hub_download
from ultralytics import YOLO
w = hf_hub_download("honi05/chess-piece-yolo", "yolov8n/best.pt")
model = YOLO(w)
results = model("board.jpg")
See the full report in the code repo github.com/Honi05/cvproject.