Computer vision engineering notes
I’m Antal Zsiros. These notes cover how I generate and review synthetic training data, evaluate computer vision models, and run OpenCV in the browser. I share the results alongside the limits of each experiment.
From 1,002 synthetic rust samples to 803 training candidates
Two review rounds, 199 exclusions and an 80.1% retention rate. Generator breakdowns, broader scene coverage and interactive mask examples before RF-DETR training.
Synthetic rust masks: GPT Image 2.5 Sunburst vs Nano Banana 2.1
Eight prompts, sixteen samples, and interactive mask overlays: where the generators help and where rust labels fail before RF-DETR training.
Training a proctoring classifier with synthetic webcam images
17,280 images generated from text prompts, a multihead model, and what a strong synthetic validation score missed in real-world testing.
Building OpenCV 5 as a Static C++ WebAssembly Library with Emscripten
A practical guide for compiling OpenCV 5 to static WebAssembly-ready C++ libraries with Emscripten, SIMD and pthreads.
Synthetic DMS Training Data Generation with Video Models
A practical experiment in generating realistic Driver Monitoring System videos together with semantic and instance segmentation masks.