Home▸AI & Machine Learning▸UK Traffic Sign Recognition: CNN vs MobileNetV2

🚦 UK Traffic Sign Recognition: CNN vs MobileNetV2

Watch a UK road sign pass through a convolutional network's conv/pool layers versus MobileNetV2's depthwise-separable blocks, and see how image corruption affects each architecture's confidence.

AI & Machine Learning3DAdvanced60 FPS

A UK road sign is captured by a virtual camera and pushed through a convolutional network's layers, rendered as a shrinking tunnel of voxel feature maps, before landing as a bar-chart prediction across seven sign classes.

🔬 What It Demonstrates

The feature maps are a real (simplified) forward pass: a Sobel edge filter on the input image, repeatedly average-pooled and coloured by activation strength, so the pipeline you see is genuinely derived from the sign's pixels — not decorative.

🎮 How to Use

Pick an architecture and a UK road sign, then drag the corruption slider to simulate rain, motion blur or low light. Run inference to watch the pulse sweep through conv/pool or depthwise-separable blocks and update the confidence bars.

💡 Did You Know?

MobileNetV2's depthwise separable convolutions cut multiply-add cost roughly 8–9× versus a standard convolution, which is why transfer-learned mobile networks can run in real time on embedded driver-assistance cameras.

⚙ Under the hood

Watch a UK road sign pass through a convolutional network's conv/pool layers versus MobileNetV2's depthwise-separable blocks, and see how image corruption affects each architecture's confidence.

artificial intelligencedeep learningconvolutional neural networkscomputer visionimage recognitionmobile netcnnThree.js

3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install