License Plate Recognition

Problem

The legacy plate pipeline was classical computer vision, and it degraded exactly where real deployments live: oblique angles, motion blur, low light, and non-standard plate formats. Accuracy in the lab did not survive the parking garage.

Approach

Replaced the classical detection and OCR stages with a deep detection-and-recognition stack. Throughput came from CUDA implementations of depth-wise convolution layers written for the in-house library.

Result

A 20% accuracy improvement over the legacy system in production, deployed as part of Sighthound’s commercial vehicle analytics products.