On-premise Computer Vision: Optimizing the 'Night Eye' Model to 4MB

Developers of the 'Night Eye' project have unveiled results from optimizing their computer vision system for local computing environments. Initially, the library occupied approximately 30 MB and required significant CPU resources; however, through library and execution optimization, the deployment size was reduced to 4.2 MB. The new version of the model fits into the processor's internal memory, eliminating the need for constant access to RAM. During testing, processing a single frame on a single CPU core takes between 8.6 and 15.8 ms, enabling efficient use on hardware-constrained devices. This approach demonstrates the feasibility of achieving high-performance machine learning without relying on cloud computing, which is critical for tasks requiring low latency and data privacy.