Estimate Crowd Count (CLIP EBC)

How it fits

Typical backends

Caveats

• Hard constraintThe weights directory referenced by model must contain exactly one `.pth` checkpoint; zero or more than one aborts startup with a count mismatch.
• AccuracyThe UInt64 total is the rounded sum of the head-density map, taken from the full-resolution map so it stays exact relative to the model, but it is integer-quantised regression rather than a verified per-head enumeration.
• AccuracyIn extremely dense crowds where heads merge the density map saturates and the total under-reports without raising.
• I/O contractThe two DensityMap outputs are emitted at the resized inference resolution (the value of model_width by model_height), not the original frame size; the count is computed before that resize so it is unaffected. Set both to 0 to run and emit at native source resolution.
• I/O contractFor a checkpoint without the dual-head density backbone, the head-density and probability outputs carry the same map.
• Parameter interactionChannel normalisation statistics are chosen automatically from the checkpoint's backbone name (CLIP-prefixed names use CLIP statistics, all others use ImageNet statistics) and the chosen pair is logged; the weights under model therefore determine preprocessing implicitly.
• FallbackA device value beginning with `cuda` silently falls back to CPU with a warning when CUDA is unavailable; CPU inference is impractical for live video.
• State lifetimeThe checkpoint and device are fixed at startup, but model_width and model_height are re-read each frame, so the inference resolution can be retuned live without a redeploy.