How to training yolov8 include slice aided hyper inference (sahi) for deploying model to android

[rahmadyd]

excuse me, i want to ask about How to training yolov8 include slice aided hyper inference (sahi) because i want to export the weight but i dont know where i put sahi code if i export that weight to tensorflowlite (i want to deploy this model to android). please tell me how it works. thanks

[glenn-jocher]

@rahmadyd to train YOLOv8 with SAHI for sliced inference and then export to TensorFlow Lite for Android deployment, you first train your model normally with YOLOv8. SAHI is used during inference, not training, so you don't need to modify your training process. After training, you can export your model to TensorFlow Lite using the export function. For integrating SAHI during inference on Android, you would handle the slicing and stitching logic in your Android application code. For detailed steps, please refer to our documentation: https://docs.ultralytics.com/guides/sahi-tiled-inference/.

[agoransson]

I've been looking for that answer for very long - I've previously understood that SAHI requires special training logic, to slice training images in the same way as images are sliced during inference. You're saying this isn't needed, "any" Yolo object-detection model should do fine while running inference with SAHI?

Thank you.

[glenn-jocher]

@agoransson that's correct—SAHI is used only during inference to handle large images by slicing them into smaller tiles. Any YOLO model trained on standard datasets should work fine with SAHI for inference without requiring special training logic.

[rygall]

If you're training models a model on unique high-resolution images to detect objects that are not in standard datasets, is there a standard built-in method to integrate SAHI into training?

[glenn-jocher]

@rygall great question! SAHI is designed for inference and doesn't require special training integration. For high-resolution training data, we recommend standard YOLO training with your full images. If you want to experiment with sliced training data augmentation, you'd need to implement that as a custom preprocessing step. The SAHI inference integration remains unchanged - see our SAHI Tiled Inference guide for implementation details.

[agoransson]

A follow-up, I've tried to find a way to run sliced inference on device. My initial tests have all failed - so I am wondering. Is there any similar technique to detect small objects available for tflite models?

[glenn-jocher]

@agoransson while there isn't a direct equivalent to SAHI (Slicing Aided Hyper Inference) specifically for TFLite models, you can explore exporting your model to TFLite with INT8 quantization for optimized performance on edge devices, as described in the TFLite - Ultralytics YOLO Docs. You might also consider optimizing at the application level on your Android device for small object detection.