// the find
AlexeyAB/yolo2_light
Light version of convolutional neural network Yolo v3 & v2 for objects detection with a minimum of dependencies (INT8-inference, BIT1-XNOR-inference)
A minimal-dependency C port of YOLOv2/v3 focused on running inference fast on constrained hardware, with INT8 quantization and 1-bit XNOR-net support. It's aimed at people who need object detection on embedded systems or edge devices where pulling in a full deep learning framework is not an option. AlexeyAB also maintains the much more active darknet fork, so this is more of a specialized offshoot than a standalone project.
The XNOR-net path is genuinely interesting — replacing multiply-accumulate with bitwise XNOR and popcount can yield 32x memory reduction and significant speedups on CPUs without vector units. INT8 inference with a calibration dataset is properly integrated, not bolted on. The dependency surface is actually minimal: CUDA/cuDNN for GPU, pthreads for CPU, optionally OpenCV — no Python, no framework. Building for CPU requires nothing but make and a C compiler, which is rare in this space.
Dead since 2019 — CUDA 8 and cuDNN 7.1 are ancient; good luck getting this to compile against anything current without patching. The XNOR training path requires you to go back to the darknet repo anyway, so the 'light' repo only handles inference, and that split is poorly documented. No CI, no tests, no mAP benchmarks in the README to tell you what you're actually giving up in accuracy with INT8 or XNOR. The directory tree is mostly hundreds of label PNG files committed to git, which is a sign no one thought hard about repo hygiene.