QuantEdge 🔬

Hardware-Aware Face Detection on Samsung GT-S7392 (2013)

Research Question: Can a 224KB neural network run face detection in real-time on a 2013 smartphone with no ML accelerator? Answer: Yes — 8.7 FPS on ARM Cortex-A9. And INT8 quantization is actually SLOWER than FP32 on this chip.

Key Finding: INT8 is 1.35x SLOWER than FP32 on ARM Cortex-A9.

Demo

BlazeFace running on ARM Cortex-A9 @ 1GHz — face detected at 92% confidence

Running BlazeFace on ARM Cortex-A9 @ 1GHz | 512MB RAM | Android 4.1.2

---

What This Project Does

Deploys a 224KB face detection model (BlazeFace) on a 2013 Samsung phone and benchmarks INT8 vs FP32 quantization on real ARM hardware.

---

Results

Stage	Hardware	Time	FPS
Laptop	Intel CPU	2.08ms	479
Phone (on-device)	ARM Cortex-A9	115ms	8.7

Key Finding: INT8 is 1.35x SLOWER than FP32 on ARM Cortex-A9. Quantization saves memory but not speed on this chip.

---

Requirements

• Python 3.10+
• Android phone with USB Debugging enabled
• Android NDK (for recompiling ARM binary)
• ADB installed

Install Python dependencies:

pip install -r requirements.txt

---

How To Run

1. Laptop Webcam Detection

python src/facedetect.py

2. Live Phone Camera Detection

Connect your Android phone via USB, enable USB debugging, then:

adb shell am start -a android.media.action.STILL_IMAGE_CAMERA
python src/live_phone.py

3. On-Device Inference (runs ON the phone)

First push the files to your phone:

adb push models/blazeface.tflite /sdcard/blazeface.tflite
adb push arm/inference_arm /data/local/tmp/inference_arm
adb push tflite_lib/libtensorflowlite_jni.so /data/local/tmp/libtensorflowlite_jni.so
adb shell chmod 777 /data/local/tmp/inference_arm
adb shell chmod 777 /data/local/tmp/libtensorflowlite_jni.so

Then run inference ON the phone:

adb shell "LD_LIBRARY_PATH=/data/local/tmp /data/local/tmp/inference_arm"

4. Recompile ARM Binary (if needed)

Download TFLite 2.4.0 AAR and extract it, then:

armv7a-linux-androideabi21-clang -o arm/inference_arm arm/inference.c \
  -I tflite_old/headers \
  -L tflite_old/jni/armeabi-v7a \
  -ltensorflowlite_jni -llog -lz -lm -ldl -landroid

---

Hardware Target

Spec	Value
Device	Samsung GT-S7392
CPU	ARM Cortex-A9 @ 1GHz
RAM	512MB
Android	4.1.2 API 16
Architecture	armeabi-v7a

---

Stack

Python · TensorFlow Lite 2.4.0 · OpenCV · ONNX · Android NDK 29 · C · ADB