[Project] My anomalib-based Live Camera Anomaly Detector

[alexriedel1]

Live defect detection in production using anomalib and webcams (and my django wrapper)

Hello!
I wanted to show you a project I built in the last months using anomalib and building a Django wrapper around it to make it useable in production settings.

It runs on Windows and Linux (and ARM64 like Nvidia Jetson) systems and is installed by only one command using uv. Its basically a Django server managing the cameras and running the anomalib training and inference.
On a Nvidia Orin NX it runs in 20ms per frame.

You can use as many cameras as the system supports. It uses synthetic anomalies for training anomlib and lets you later refine the normalization and threshold values.

1. Start

start.mp4

2. Adding a new project (using 2 cameras) and Setting the camera properties

create_project-1.mp4

3. Capturing Images of the desired area

collect_images.mp4

4. Organizing the Training dataset

show_images.mp4

5. Training a model

train_model.mp4

6. Performing Inference (it lets you set the image and pixel min max and thresholds, so you can adjust everything in production)

inference.mp4

I guess thats what Intel Geti actually does, but is a bit harder to install and i only needed these basic functions so untrained users can run it!

Right now ist not open-sourced but maybe it will be soon. If you have any questions, Tips or whatever just leave a comment :)

[samet-akcay]

@alexriedel1, this is great! Thanks for sharing!
FYI @Radwan-Ibrahim @ashwinvaidya17 @MarkRedeman @maxxgx @camiloHimura

We are working on a lightweight version of Geti, tailored for Anomalib, and I can see nice features in your project we could be inspired by.
https://github.com/open-edge-platform/anomalib/tree/feature/geti-inspect