Raspberry Pi deployment in Computer Vision - ML Experiment: Train & Evaluate

Experiment - Raspberry Pi deployment

Problem:You have a computer vision model trained on a desktop that detects objects in images. Now you want to deploy this model on a Raspberry Pi to run real-time object detection using the Pi's camera.

Current Metrics:On desktop: accuracy 90%, inference time 50ms per image. On Raspberry Pi: model runs but inference time is 2 seconds per image, making real-time detection impossible.

Issue:The model is too large and slow for the Raspberry Pi hardware, causing slow inference and poor real-time performance.

Your Task

Optimize the computer vision model and deployment code to achieve inference time under 300ms per image on Raspberry Pi while maintaining accuracy above 85%.

You cannot retrain the model from scratch on the Pi.

You must use the existing trained model weights.

You can modify the model architecture, quantize the model, or optimize code for the Pi.

Hint 1

Hint 2

Hint 3

Hint 4

Solution

Computer Vision

import cv2
import numpy as np
import tflite_runtime.interpreter as tflite

# Load TFLite model and allocate tensors
interpreter = tflite.Interpreter(model_path='model_quantized.tflite')
interpreter.allocate_tensors()

# Get input and output tensors details
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

# Initialize camera
cap = cv2.VideoCapture(0)

while True:
    ret, frame = cap.read()
    if not ret:
        break
    # Resize frame to model input size
    input_shape = input_details[0]['shape']
    input_height, input_width = input_shape[1], input_shape[2]
    img = cv2.resize(frame, (input_width, input_height))
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img = img.astype(np.float32) / 255.0
    img = np.expand_dims(img, axis=0)

    # Set tensor
    interpreter.set_tensor(input_details[0]['index'], img)
    interpreter.invoke()

    # Get output
    output_data = interpreter.get_tensor(output_details[0]['index'])

    # Simple threshold to detect object presence
    if output_data[0][0] > 0.5:
        cv2.putText(frame, 'Object Detected', (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 1, (0,255,0), 2)

    cv2.imshow('Raspberry Pi Object Detection', frame)
    if cv2.waitKey(1) & 0xFF == ord('q'):
        break

cap.release()
cv2.destroyAllWindows()

Converted the original model to TensorFlow Lite format with 8-bit quantization to reduce size and speed up inference.

Used tflite_runtime interpreter optimized for Raspberry Pi instead of full TensorFlow.

Reduced input image size to match model input to lower computation.

Implemented real-time video capture and inference loop with efficient pre-processing.

Results Interpretation

Before: Inference time 2 seconds per image, accuracy 90%.

After: Inference time 250ms per image, accuracy 87%.

Model quantization and using lightweight runtime libraries can greatly improve inference speed on limited hardware like Raspberry Pi, with only a small accuracy trade-off.

Bonus Experiment

Try deploying the model using a different lightweight architecture like MobileNetV2 and compare inference speed and accuracy on Raspberry Pi.

💡 Hint

MobileNetV2 is designed for mobile devices and can be converted to TensorFlow Lite with quantization for fast inference.