Performance Tuning

Balancing Latency vs. Throughput is the key challenge in computer vision.

Strategies

🏎️ "I need Speed" (Low Latency)

Use this profile if you are demoing the system and need immediate feedback (e.g., <100ms lag).

1. Lower BATCH_WAIT_MS: Set to 10ms.
   • Effect: GPU runs more often with smaller batches. Higher overhead but quicker results.
2. Increase PROCESS_FPS: Set to 15+.
   • Effect: Smoother movements, provided the GPU can keep up.

🏭 "I need Scale" (High Throughput)

Use this profile if you are deploying to a stable with 20+ cameras on limited hardware.

1. Increase BATCH_WAIT_MS: Set to 100ms.
   • Effect: Allows the system to group up to 32 images into one huge batch. Extremely efficient utilization of CUDA cores.
2. Lower PROCESS_FPS: Set to 2 or 3.
   • Effect: Motion looks "choppy" to a human, but the AI still detects "Standing/Lying" perfectly fine. Saves massive CPU/Decodes resources.

Hardware Sizing

Scenario	GPU VRAM	Recommended Cameras	Config Profile
Dev / Demo	4 GB	1 - 4	Low Latency
Small Stable	8 GB	5 - 12	Balanced
Commercial	24 GB (A10G)	30 - 50	High Throughput