A reproducible workflow for diagnosing YOLO26-to-NCNN deployment inconsistencies on resource-constrained Android...

Area of Science:

• Computer Vision
• Machine Learning Deployment
• Mobile Computing

Background:

• Edge devices demand efficient conversion of desktop AI models for mobile deployment.
• Resource-constrained Android devices present unique challenges for running complex models like YOLOv2.
• Silent deployment errors can significantly degrade model performance and reliability on mobile platforms.

Purpose of the Study:

• To present a reproducible workflow for diagnosing and resolving YOLOv2-to-NCNN conversion issues on Android.
• To address silent deployment errors impacting model accuracy and functionality in Android runtimes.
• To validate the workflow across various YOLOv2 asset types and on physical Android devices.

Main Methods:

• Developed a five-stage fault localization process from symptom identification to artifact repair.
• Integrated export, Android asset management, Java-JNI-C++ inference, and logging.
• Utilized frozen-image regression and CPU-path evidence for validation.

Main Results:

• Successfully localized and repaired common silent errors including letterbox, activation, coordinate-semantic, and layout faults.
• Demonstrated the workflow's effectiveness on a 10-class SafeHat PPE detection task.
• Validated the approach across detection, segmentation, pose, classification, and oriented-box YOLOv2 assets.

Conclusions:

• The presented YOLOv2-to-NCNN diagnosis workflow enables reliable deployment of AI models on Android.
• The five-stage fault localization method effectively identifies and rectifies silent deployment errors.
• The validation package provides a robust resource for ensuring accurate AI model inference on mobile devices.