Leveraging the power of C++ for efficient machine learning on embedded devices
summary
Shows how C++ with TensorFlow Lite runs image models on a Raspberry Pi Model B (4-core, 4 GB) using MobileNet for image classification and ResNet50 transfer learning for rock–paper–scissors; models are ~23 MB, inference ~34 ms on still images and ~60 ms live, memory under 100 MB. Trained on a small set (125 images/class) vs a large set (≈90,964 images/class), yielding ~65% vs >90% overall accuracy; a real‑time camera hand‑gesture game is implemented and works, though scissors remain error‑prone.
AI-written summary · about
chapters (16)
- 0:00Introduction and motivation
- 3:22Embedded devices characteristics
- 4:52Why use C++ for ML
- 7:26Image classification basics
- 9:13Raspberry Pi setup for ML
- 12:25C++ code for image classification
- 16:49Image classification demo
- 20:08Analysis of results
- 22:43Performance comparison with Python
- 23:07Hand gesture recognition task
- 24:34Transfer learning approach
- 27:49Real-time image classification
- 31:39Rock Paper Scissors game demo
- 33:10Summary and future work
- 34:35Conclusion and data importance
- 36:22Q&A session
related talks
48mLeveraging the Power of C++ for Efficient Machine Learning on Embedded Devices
27mMachine Learning on the edge using TensorFlow Lite
17mBuild a smart IoT device with TensorFlow Lite and Raspberry Pi
46mTensorFlow for Mobile Poets
34mTensorFlow Lite
43mTensorFlow for Mobile Developers
32mAI for Mobile and IoT Devices: TensorFlow Lite
42mWhat's New with TensorFlow?