What it does

This repo is a port of YoloV3 and tiny-YoloV3 to Intel’s OpenVINO toolkit, with explicit support for Raspberry Pi 3, Ubuntu laptops, and the Neural Compute Stick (NCS/NCS2). It includes both Python and C++ inference pipelines, plus scripts that can distribute work across multiple NCS2 sticks for a modest speed boost.

The interesting bit

The author treats the NCS2 almost like a GPU you plug in over USB: one stick gets you ~30 FPS with tiny-YoloV3 on a Core i7, and four sticks push full YoloV3 to 13 FPS. The README is refreshingly honest about the trade-offs, ranking its own models against MobileNet-SSD on speed, accuracy, and detection distance.

Key highlights

• Python and C++ async inference demos with USB camera or MP4 input
• Multi-stick support via -numncs flag (up to 4 NCS2 devices)
• Pre- and post-processing bug fixes in March 2019 that improved accuracy
• Custom training pipeline added for proprietary datasets
• Targets OpenVINO 2019 R1.0.1, TensorFlow 1.12, OpenCV 4.1.0-openvino

Caveats

• Last meaningful update was April 2019; OpenVINO and TensorFlow have moved on significantly
• Full YoloV3 on Raspberry Pi 3 is noted as “pretty slow” due to ARM vs. Core i7 performance gap
• The C++ version lives in a separate cpp/ directory with its own build requirements

Verdict

Worth a look if you’re maintaining legacy NCS2 hardware or need a concrete reference for OpenVINO model conversion. Skip it if you’re starting fresh—modern YOLO variants and newer Intel toolkits have simpler, faster paths to deployment.