Hand Gesture Recognition — YOLO26n

Fine-tuned YOLO26n for 12-class hand gesture recognition, trained on the HaGRID dataset.

Model Details

Classes

0: like      1: palm      2: fist       3: peace
4: ok        5: one       6: three      7: four
8: call      9: point    10: grip      11: middle_finger

Usage

from ultralytics import YOLO

model = YOLO("yolo26n_hand-gesture_12c_v1.pt")
results = model("hand.jpg")
results[0].show()

Training Parameters

• imgsz: 800
• batch: 48
• optimizer: auto (AdamW)
• device: CUDA (NVIDIA RTX 5070 Ti, 16 GB)
• cache: disk
• workers: 12
• patience: 0

Dataset

Trained on a subset of HaGRID (Hand Gesture Recognition Image Dataset), containing ~36,000 images across 12 gesture classes.

Performance Notes

The model converged around epoch 84 (mAP50-95=0.862) and remained stable through epoch 100. Suitable for real-time inference on consumer GPUs.