ehcalabres/chess-pieces-rfdetr-small

RF-DETR Small — chess piece detector (12 classes)

Fine-tuned RFDETRSmall (resolution 560, COCO-pretrained) for chess piece detection from camera images. Trained for 50 epochs (batch 8 × grad_accum 2, lr 1e-4, EMA) with rfdetr.

• Data: ChessReD2K bbox subset (CC BY-NC-SA 4.0) + acapitani/chesspiece-detection-yolo (MIT), unified 12-class COCO — train 3,220 imgs / ~67.7k boxes, valid 760 / ~16.1k, test 306 / 6.3k. Dataset: https://huggingface.co/datasets/ehcalabres/chess-rfdetr-coco
• Classes: white-pawn, white-rook, white-knight, white-bishop, white-queen, white-king, black-pawn, black-rook, black-knight, black-bishop, black-queen, black-king
• Final test metrics (epoch 50): mAP@50 = 1.000, mAP@50:95 = 0.941, mAP@75 = 0.998, F1 = 0.9999
• Best val EMA mAP@50:95: 0.906 (epoch 49); full curve in metrics.csv and the trackio Space
• Usage:

  from rfdetr import RFDETRSmall
  model = RFDETRSmall.from_checkpoint("checkpoint_best_ema.pth")
  detections = model.predict("board_photo.jpg", threshold=0.5)
  

Non-commercial use notice: trained partly on ChessReD (CC BY-NC-SA 4.0) — non-commercial use only.