SIMD-accelerated computer vision on a $2 microcontroller
Summary
The article describes a SIMD-accelerated FAST feature detector implemented on the ESP32-S3, achieving about 2x performance improvements (roughly 6 ms per 320x240 frame) and reaching 11.2 MP/s throughput for real-time vision. It covers hardware constraints, clever tricks to map unsigned pixel values to a signed range for SIMD, and provides code fragments and insights into assembly and vector operations. The work demonstrates impressive low-cost computer vision on a $2 microcontroller with source code available on GitHub.