Skip to content

References: Building a Real-Time Object Detection Pipeline

  1. Quantization (signal processing) - Wikipedia - Explains how quantization maps a large or continuous set of values to a smaller set through rounding, and why it underlies lossy compression. Foundational math behind this chapter's model quantization from 32-bit to 8-bit numbers.

  2. Transfer learning - Wikipedia - Describes how knowledge learned on one task, like recognizing cars, can boost performance on a related task, like recognizing trucks, without training from scratch. Matches the chapter's use of a pretrained model to add a custom object class.

  3. Edge computing - Wikipedia - Covers running computation near the data source rather than in a distant data center, and the resulting latency, bandwidth, and privacy benefits. Directly grounds this chapter's definition of edge AI on the Raspberry Pi 5.

  4. Efficient Processing of Deep Neural Networks (Synthesis Lectures on Computer Architecture) - Vivienne Sze, Yu-Hsin Chen, Tien-Ju Yang, and Joel S. Emer - Morgan & Claypool Publishers - Credited with establishing the standard energy/latency/accuracy tradeoff framework for evaluating model quantization and compression, the same size-speed-accuracy tradeoff this chapter's Chart.js explorer visualizes.

  5. Deep Learning for Coders with fastai and PyTorch (1st Edition) - Jeremy Howard and Sylvain Gugger - O'Reilly Media - Credits Howard's fastai library and ULMFiT-derived discriminative fine-tuning and gradual unfreezing techniques for popularizing practical transfer learning, now the standard way of retraining only a pretrained model's final layers.

  6. Post-training quantization - Google / LiteRT (TensorFlow Lite) Documentation - Official guide to dynamic-range, full-integer, and float16 quantization techniques for shrinking a trained model for edge deployment, with a decision framework matching this chapter's compression-tradeoff discussion.

  7. ONNX Get Started - ONNX Official Documentation - Outlines the workflow for building a model, exporting it to ONNX format, and running inference across different frameworks. Illustrates why ONNX matters as a shared, tool-independent model format for edge deployment.

  8. What is Non-Maximum Suppression? - GeeksforGeeks - Explains greedy and soft NMS with the IoU-based overlap calculation and Python/OpenCV code, matching this chapter's pseudocode for suppressing duplicate overlapping bounding boxes. Covers use in YOLO and other detectors.

  9. Transfer learning and fine-tuning - TensorFlow Official Tutorial - Walks through freezing a pretrained MobileNetV2 base, adding a new classification head, and later unfreezing top layers to fine-tune. Concrete code companion to this chapter's transfer learning workflow diagram.

  10. AI software - Raspberry Pi Documentation - Official guide to installing Hailo NPU drivers and running pretrained YOLOv5/v6/v8 and YOLOX object detection models through rpicam-apps on the Pi 5. Matches this chapter's camera-to-model pipeline on AI HAT+ hardware.