YOLOv10-LiteRT on Android for Object Detection

Real-time on-device object detection on Android using YOLOv10 converted to LiteRT (.tflite) via Google AI Edge.

Pipeline

How it works

The conversion goes through a few steps that aren't obvious from the outside. Ultralytics doesn't export directly to TFLite — it first converts the PyTorch model to ONNX, then uses onnx2tf to produce a TensorFlow SavedModel, and finally converts that to .tflite. The yolov10_litert.py script handles the full chain with a single --export flag.

Once on Android, inference runs through the LiteRT runtime (formerly TFLite) using the Detector class, which:

• Preprocesses camera frames to 640×640 float32 tensors
• Reads the model output tensor of shape [1, 300, 6] — 300 candidate detections, each with (x_min, y_min, x_max, y_max, confidence, class_id) in normalized coordinates
• Scales the box coordinates back to screen dimensions
• Filters by confidence threshold and passes the results to OverlayView

OverlayView draws bounding boxes and labels directly on the camera preview using Android's Canvas API, giving you a live overlay with no perceptible lag on modern devices.

The labelmap is extracted from the metadata.yaml that Ultralytics writes alongside the exported model and saved as a labels.json file that the Android app loads at startup.

Once the model is exported, you can also visualize its full neural network architecture interactively using Google AI Edge's Model Explorer:

import model_explorer
model_explorer.visualize("yolov10n_saved_model/yolov10n_float32.tflite")

This opens a browser-based graph view where you can inspect every layer, see tensor shapes flowing through the network, and understand how YOLOv10's head produces the [1, 300, 6] detection output. Useful for debugging quantization issues or just understanding what's inside the model.

Demo Output

How to run

Python — export the model

pip install -r requirements.txt

# Export YOLOv10n to TFLite (downloads weights automatically)
python yolov10_litert.py --export --model yolov10n.pt

# Test inference on an image
python yolov10_litert.py --input test_image.jpg --tflite yolov10n_saved_model/yolov10n_float32.tflite

# Test inference on a video
python yolov10_litert.py --input test_video.mp4 --tflite yolov10n_saved_model/yolov10n_float32.tflite --confidence 0.4

Android — deploy the model

1. Copy the exported .tflite file and labels.json into Android_App/app/src/main/assets/
2. Open Android_App in Android Studio
3. Build and run on a physical device (API 24+)
4. Grant camera permission on first launch

The app runs everything on-device — no network calls, no server.

Resources

1. 
2. Medium Blog — step-by-step walkthrough
3. YOLOv10 docs by Ultralytics
4. Google AI Edge LiteRT

Acknowledgment

This project was developed during Google's ML Developer Programs AI Sprint. Thanks to the MLDP team for providing Google Cloud credits to support this project.