Audley Weird: You Only Look Once
With the release of YOLO26 recently, correlo was eager to get our hands dirty and put its new capabilities to the test.
For the uninitiated, YOLO stands for "You Only Look Once". It is a state-of-the-art, real-time object detection model that has become our absolute go-to for a massive variety of computer vision applications.
But what exactly is object detection? Simply put, it’s a computer vision technique that allows software to identify and locate specific objects within an image or video - essentially teaching a computer to "see" and comprehend the world like a human does.
To test out YOLO26, we needed a subject. Enter Audley Weir, a favourite chill-out spot located deep in the Royal National Park in New South Wales.
It's famous for flooding and cutting off traffic, which might be why Transport for NSW operates public traffic cameras there. We decided to point our new AI eyes at this specific camera, giving birth to our internal testing project: Audley Weird.
Deployment
YOLO26 is a new version for 2026 which is heavily optimised and built for edge computing - with the aim to perform AI inference on edge devices such as cameras, drones, and other low-powered devices in the field.
However, out at the Royal National Park, it's really hot outside. Rather than sitting around in the hot Sydney summer sun typing code and waiting for cars to pass, we realised that YOLO26’s lightweight, fast-inference characteristics also make it a fantastic candidate for AWS Lambda.
AWS Lambda is a serverless computing service. It is ephemeral, highly scalable, and can be cheap - because you only pay for the exact milliseconds of compute time you use. Since we didn't need to run a heavy server 24/7 full of precious GPUs to process intermittent images, Lambda was the perfect environment for this kind of rigorous, event-driven testing.
Usually, we train our models for custom inference to detect highly specific objects. However, for Audley Weird, we were more interested in getting some initial benchmarks on the new YOLO model, and so we utilised YOLO's out-of-the-box object classes. YOLO comes pre-trained on the COCO (Common Objects in Context) dataset, which has robust native recognition for vehicles such as cars, trucks, motorcycles, and buses.
With the backend humming along on Lambda, we needed a front-end to visualize the data.
Instead of a boring stream of text, we built out the Audley Weird interface and introduced tight segmentation masks to 'chop out' the vehicles as they appear across the weir in real time.
Constraints
Building this wasn't without its hurdles. Working with this particular camera feed came with some challenges:
Image Quality: The public NSW traffic cameras provide notably low-resolution and poor-quality images. This is probably a deliberate design choice by Transport for NSW to retain public privacy, ensuring that license plates and drivers' faces cannot be recognised. While great for privacy, blurry pixels make the object detection model work a lot harder!
Low Frame Rate: We aren't working with a live video feed. The camera only updates with one static image every 20 seconds, meaning our architecture had to be optimised for discrete, scheduled inferences rather than a continuous stream.
The Output
You can see the final result of our experiment in action over at www.audleyweird.com.
Head over to check out the live vehicle detections, view our real-time inference speeds, and play around with the physics of the site (yes, you can throw the detected cars around!).
At correlo, the vibe is different. We hope Audley Weird demonstrates our willingness to take on absolutely any brief, no matter how obscure, challenging, or downright weird it might be.
Got a complex computer vision problem, or just a fun idea you want to bring to life?
Please send your puzzles through to hello@correlo.com.au and let's chat!
