PeekingDuck

PeekingDuck's core idea is pipeline-as-config:

``yaml nodes: - input.visual: source: webcam - model.yolo - draw.bbox - output.screen``

Once launched, that config becomes a complete real-time pipeline: webcam → YOLO detection → bounding boxes → display. The framework ships with 50+ built-in nodes spanning input (video, webcam, images), models (YOLO, HRNet, PoseNet, etc.), post-processing (tracking, counting, ROI filtering), and output (screen, file, message queue).

Use cases: intelligent surveillance, footfall analysis, retail customer behavior, sports pose analysis, safety-compliance checks (masks, hard hats).

PeekingDuck takes an unusual path in the CV-tooling ecosystem: not chasing SOTA models, but building "the simplest framework you can actually run in production."

The industry already has powerful CV tools like OpenCV, Detectron2, and MMDetection, but they are too high-friction for small businesses, students, and non-ML developers. PeekingDuck aims to let a developer who "knows Python but not deep learning" stand up a production-grade CV application in 30 minutes.

Under the hood it wraps PyTorch, TensorFlow, and other frameworks behind a simple interface. Performance is not best-in-class, but the deployment, debugging, and maintenance overhead is far lower than rolling your own pipeline.

PeekingDuck is another product of AISG's "open-source tooling strategy" — like TagUI, it's positioned to lower the barrier to AI deployment.

Across the seven transmission levers:

• Lever 3 (Industry Adoption): brings CV technology within reach of local SMEs
• Lever 6 (Diplomacy): an open-source project serving as a vehicle for Singaporean AI exports

Take: PeekingDuck isn't AISG's most famous project, but it embodies AISG's engineering philosophy: build "good enough" tools rather than SOTA tools. That pragmatic line is reasonable for a small market like Singapore — don't try to out-scale OpenCV or Meta, but differentiate on ease of use.