What it does This repo is a collection of ~20 standalone C applications that run pre-trained deep learning models for face detection, pose estimation, depth estimation, style transfer, and a half-dozen other vision tasks. Each demo uses OpenGL ES for rendering and either TensorFlow Lite GPU Delegate or TensorRT for inference. Targets are Linux PC, NVIDIA Jetson Nano, and Raspberry Pi.

The interesting bit The author cross-compiled the entire zoo for ARM, wrote per-platform Makefiles, and even warns that GPU Delegate on Raspberry Pi performs worse than CPU — then documents the XNNPACK fallback. That’s the kind of honesty that saves you an afternoon.

Key highlights

• 20+ working demos: Blazeface, DBFace, FaceMesh, 3D hand pose, DenseDepth, AnimeGAN, MIRNet low-light enhancement, and others
• Dual backend support: TensorFlow Lite GPU Delegate v2 or TensorRT (noted per-app where available)
• Build scripts for x86_64, aarch64 (Jetson/Pi 64-bit), and armv7l (Pi 32-bit)
• VSYNC toggle documented for Jetson to trade tearing for frame rate
• CI badge shows TFLite 2.7 builds are being exercised

Caveats

• Build process is manual and old-school: install Bazel 3.1.0, compile TensorFlow Lite from source, scp libraries to target, then make
• README cuts off mid-sentence during the armv7l section, suggesting it may be stale or truncated
• Several “TensorRT port is HERE” links imply the main app is TFLite-first with TRT as secondary

Verdict Worth cloning if you need a working baseline for edge inference on Pi-class hardware and prefer C over Python. Skip it if you want pip-install convenience or a unified framework — this is a craftsperson’s collection of individual demos, not a product.