But why bother with image recognition? And pay a heavy latency penalty which makes it difficult to implement PID-style controls on player's aim? Its a game of milliseconds and opencv is comparably slow, as is the I/O involved.

Far easier to do a MiTM on the network traffic and recreate state in another process.

Still easier to clone memory pages used by the target app, or use hardware that allows dma.