🚦 UK Traffic Sign Recognition: CNN vs MobileNetV2

About this simulation

A UK road sign is captured, then pushed forward through a convolutional network. The visualisation performs a real (simplified) forward pass: a Sobel edge filter runs on the input image, and each subsequent layer is a genuine average-pooled downsample of that edge map, coloured by activation strength — so the shrinking "tunnel" of voxel grids you see really does derive from the pixels of the sign you picked.

Custom CNN vs MobileNetV2

• Custom CNN — ordinary 3×3 convolutions followed by 2×2 max-pooling, trained from scratch on the 43-class GTSRB-style dataset. Every filter must be learned from zero, so it needs more training data to reach high accuracy and is more sensitive to noisy or corrupted images.
• MobileNetV2 (transfer learning) — replaces standard convolutions with depthwise separable blocks: a depthwise 3×3 filter per channel, then a pointwise 1×1 convolution to mix channels. Its early layers are pretrained on ImageNet, so it starts from rich general-purpose features and needs far fewer multiply-adds per layer, while typically staying more confident under distortion.

Controls

• Architecture — switch the network pipeline between the custom CNN and MobileNetV2's inverted-residual blocks.
• Road sign — choose which UK sign is fed into the camera plane.
• Image corruption — simulates rain, motion blur or low light on the captured photo; watch confidence fall and, past a threshold, the prediction can flip to the wrong class.
• Run inference / Auto-run — fires a pulse that sweeps through the layers and updates the output bars, mimicking a forward pass.

Did you know?

MobileNetV2's depthwise separable convolutions cut the multiply-add cost of a standard 3×3 convolution by roughly 8–9× for typical channel counts — a key reason it runs comfortably on the embedded processors found in driver-assistance cameras.