Personal project · On-device computer vision

PelagiVex — Computer Vision Prototypes

Two related explorations: underwater gate detection and a separate native macOS camera segmentation preview, presented as a visual system overview.

← All projects View source on GitHub

Architecture

Two independent paths, one local project.

A conceptual illustration of the components, not live model output.

01 / UNDERWATER GATE

01Image or sampled frameBrowser input and resizing
02Java + ONNX RuntimeFine-tuned RF-DETR Nano
03Gate detectionsBoxes on the analysed frame

02 / NATIVE CAMERA

01Camera frameSwiftUI + AVFoundation
02Core ML segmentationSeparate YOLO11 model
03Masks + approximate depthLocal depth worker; uncalibrated

MacBook Pro · local execution · no server inference

Underwater gate detection

The research prototype connects a bilingual browser interface to a Java inference application. Images and sampled video or camera frames are prepared by the browser and passed to an ONNX Runtime implementation of a fine-tuned RF-DETR Nano model. Video is analysed as individual frames rather than uploaded as a complete video.

A separate native camera preview

An independent SwiftUI and AVFoundation app uses Core ML for on-device instance segmentation. Its general-purpose YOLO11 segmentation model is separate from the underwater gate detector. A local depth worker provides approximate, uncalibrated depth estimates for segmented objects.

Where it runs

The prototypes run locally on my MacBook Pro. This portfolio page is a static explanation of the architecture; it does not run either model or provide a live detection demo on the server.

Outputs & learning

A browser-to-inference workflow for individual image and video frames.
An independent native camera preview combining segmentation and approximate depth.
A visual explanation of how interfaces, models and outputs fit together.

Scope & limitations

The supplied dataset has near-duplicate frames across splits; its test results do not establish generalisation to new scenes.
The native model recognises pretrained general-purpose classes, not underwater gates; depth estimates are uncalibrated and can lag.
Research and development prototypes only: no vehicle control, production reliability or real-world accuracy claim.
The gate model artifact is not distributed with the public repository.

Evidence basis

Local PelagiVex source and project documentation, including the native preview README and model card; no new model evaluation was run for this portfolio page.

View source on GitHub ↗