MemryX Inc. ads

Real-time bakery item counting on the MemryX Inc. MX3 accelerator card 🍦 In food production lines, counting sounds simple until it isn’t. Items move fast, overlap on conveyors, change orientation, and sometimes partially occlude each other. Whether it’s ice-cream cones on a belt or cream being filled into nests, maintaining accurate counts in real time is critical for quality control and throughput. We built a bakery analytics pipeline that tracks and counts items directly on the edge. Here’s how it works: ✅ Real-time detection of bakery items (ice-cream cones, nests, etc.) using a YOLO26 model. ✅ Tracking across frames to ensure consistent counting even with occlusions. ✅ Items counting with line counter using Shapely. ✅ Live count visualization for production monitoring ⚙️ Pipeline flow: YOLO26 → ONNX → DFP compilation → inference on MX3 → tracking & counting analytics 👉 The outcome: A reliable, edge-based counting system that keeps up with high-speed production lines, delivering accurate counts without adding system complexity or cloud dependency. 🔗 Code ➡️https://lnkd.in/gpPhaS8N

The MemryX May newsletter is out. → SDK v2.2: up to 4x FPS on YOLO models, expanded ONNX/Keras support, unified C++/Python API → YOLO26 runs out of the box on MX3 — 40+ variants, no extra export steps → Real-time traffic analysis example on MX3 → Partner spotlight: baggage monitoring with Arbor Solution and DeepX → MemryX at Automate 2026 — Chicago, June 22–25 Scan the QR code to subscribe.

How accurate does your AI model actually need to be? Accuracy means different things depending on the task. A cancer detector needs high recall. A car counter doesn't need to catch every partially-obscured vehicle. Optimizing blindly for one number often misses the point. The latest MemryX Dev Blog post breaks down what accuracy really means in practice — the metrics that matter, what affects model performance, and how MX3 preserves accuracy when you deploy at the edge. Read it here: https://lnkd.in/gixjCy8f #EdgeAI #MemryX #MachineLearning #ComputerVision

Real-time fire detection using YOLO11 on the MemryX Inc. MX3 accelerator card 🔥🏠 Fire accidents can escalate within seconds. Early detection is critical, whether it’s inside a kitchen, living room, garage, or outside near entrances and storage areas. This real-time fire detection system is running entirely at the edge using the MX3 AI accelerator card. How this demo is built: ✅ Compiled a custom-trained YOLO11 fire detection model into DFP format for inference. ✅ Detected flames and fire instances from both indoor and outdoor camera feeds. ✅ Displayed bounding boxes directly on video frames. Future work: 💡 Integration to trigger alarms or notifications when fire is detected. ⚡ Pipeline flow: YOLO PyTorch → ONNX → DFP compilation → Inference on MX3 → Fire detection Explore the complete code example ➡️ https://lnkd.in/g4C5QPaY

What comes after your first Arduino or Raspberry Pi project? For many makers and engineers, the next step is computer vision — letting your hardware respond to what it sees. We published an intro guide to integrating computer vision into hardware projects using the MemryX MX3 AI Accelerator. Whether you're on a Raspberry Pi or controlling things from an Arduino, this series shows you how to bring real-time AI into your builds. Read the blog: https://lnkd.in/gud_DFGC #ComputerVision #EmbeddedSystems #Arduino #RaspberryPi #AIHardware #Makers #EdgeAI #MemryX

MemryX Inc. released SDK 2.2 with YOLO26 support 🚀 One of the hidden costs in deploying object detection pipelines isn’t just the model; it’s everything around it. Post-processing, especially non-maximum suppression (NMS), adds latency and complexity, and quickly becomes a bottleneck when scaling to real-time, multi-stream workloads. With SDK 2.2, a lot of the improvement comes from the compiler. Newer YOLO architectures, especially those using attention blocks, now map much more cleanly to MX3. In many cases, models that previously required optimization can now run as-is. 👉 The practical part: Take your ONNX model and just run it. No extra export paths. No model changes. Pipeline: YOLO26 → ONNX → DFP compilation → NMS-free inference on MX3 → downstream analytics Note: This demo uses a pretrained YOLO26 model (COCO). Custom models and more industrial use cases coming soon. 🔗 Test YOLO26 on MX3 ➡️ https://lnkd.in/gkSEj_uk