Hi everyone, I'm a final-year Electronics Engineering student, and my team is working on our major project. I'd really appreciate feedback from people who have experience in computer vision, embedded AI, or machine learning deployment. Our goal is to build a **portable, offline sign language recognition system** that runs entirely on a **Raspberry Pi 5** without any cloud dependency. ### Current system design ur proposed pipeline is: * Raspberry Pi 4 + Camera Module 3 for live video capture * MediaPipe Hands to extract 21 hand landmarks (63 features) * Landmark normalization to reduce the effects of hand size, position, and camera distance * Lightweight classifier running with TensorFlow Lite * INT8 quantization for faster inference on Raspberry Pi * OLED display for text output * Offline Text-to-Speech for voice output The initial target is to recognize the **26 ASL alphabet gestures**, with plans to expand later. ### Why we chose landmark-based recognition Instead of feeding raw RGB images into a CNN, we're using MediaPipe landmarks because they: * Greatly reduce computational cost * Require much less memory * Preserve user privacy * Are better suited for real-time inference on edge devices I'd really value your opinions on the following: 1. **Model selection:** Since the input is only a 63-dimensional landmark vector, would you recommend an MLP, 1D CNN, GRU, LSTM, Transformer, or another architecture? What would you choose if the priority is real-time inference on a Raspb…

Full article content could not be extracted automatically. Read the original below.