Open-source tools for computer vision and perception.
| Project | Description |
|---|---|
| modern-yolonas | Clean re-implementation of YOLO-NAS in modern PyTorch. |
| pexel-downloader | Download photos and videos from Pexels. |
Python, PyTorch, ONNX, OpenVINO, CoreML, TensorRT
UrbanOmniDetect-2 (monocular 3D vehicle boxes + top-down map) measured on a laptop: OpenVINO INT8 on an Intel iGPU, CPU, and TensorRT, with parity against PyTorch
Companion code for the condados.ai Fundamentals track, measured rather than asserted. Today: the Edge Detection unit — gradient operators, LoG zero-crossings, Canny's NMS and hysteresis taken apart, and the structure tensor behind Harris, Shi-Tomasi and Forstner.
How many good parts does PatchCore need to beat a VLM that has seen none? Jev-Omni and Gemma 4 12B zero-shot vs PatchCore on VisA.
What Jev-Omni's decision head buys: three readouts of one Gemma 4 12B forward pass on DecisionBench, BLINK and MMStar, with calibration
Marigold V2 is a rank-128 LoRA on a frozen 20.4B backbone, so the backbone is swappable. What that costs, measured on one A40.
A background-segmentation video filter for OBS Studio in C++ on OpenVINO, with the plumbing measured stage by stage
Where repetition counting breaks in a pose pipeline: jitter with memory, the filter that earns its place, and the state machine that does the work. Companion code for condados.ai.
Reproducing Shah et al.'s webcam pupil-diameter model over the full EyeDentify dataset, and converting it to run in a browser. Companion to a CondadosAI article.