Muhammad Ali Taha

Summary

What I did

• Trained a YOLO11m ball detector with motion-blur augmentation; benchmarked and shipped an RF-DETR-Large alternative (95.4% mAP@50, DINOv2 backbone, Hungarian-matching set-prediction loss) that cut per-clip latency 10× and boosted small-object recall.
• Eliminated manual pitch calibration by training a YOLO11m-pose 10-keypoint pitch detector from scratch wired Lucas-Kanade optical flow across he 4 corners so the pixel-to-metre homography stays locked through camera zoom, pan, and mid-over cuts.
• Diagnosed a failure mode in the published CricShotNet paper (pretrained model scored ~5% on real broadcast frames vs. claimed 89%), replaced the broken YOLO segmentation stage with Meta's SAM 2 box-prompted oundation model, and retrained an EfficientNetV2-S + GRU classifier from scratch on 11,214 clips.
• Trained a model on different Mushaf Ayah markers.
• Performed model fine-tuning.

Results

• Replaced manual labeling by developing a generic Ayah marker detection solution.
• Enabled users to play audio from specific Ayahs through automated marker identification.
• Achieved 95.4% mAP@50 with RF-DETR-Large alternative.
• Retrained classifier on 11,214 clips.
• Achieved more than 95% recall and accuracy for Ayah marker detection.
• Achieved more than 85% mAP@50-95 across different Mushaf styles.