Detect Objects (DEIMv2)

How it fits

Typical backends

Caveats

• Hard constraintThe configured model must name a HuggingFace repo holding a DEIMv2 COCO checkpoint; an empty value aborts startup. The ultra-light HGNetv2 sizes are self-contained, while the DINOv3-backed sizes are heavier and more accurate.
• Parameter interactionAn emitted class id is the zero-based COCO index the checkpoint was trained on; the class filter and the per-class confidence overrides both key off that same index, and a class without an override falls back to the global confidence minimum.
• AccuracyThe square inference resolution should match the value the checkpoint was trained at (320 for the Atto size, 416 for Femto, 640 for the larger sizes); a mismatch lowers accuracy without raising an error.
• LimitationThis is a detection-only model with no instance-mask output arm; for masks use a segmentation component instead.
• FallbackA frame that decodes empty or errors during inference is logged and yields the safe empty value for the resolved arm (an empty list, a zero UInt64 count, or the untouched frame on the overlay arm), which is indistinguishable from a genuinely empty scene; only startup misconfiguration raises.
• State lifetimeAll configuration keys and the loaded weights are captured once at startup; runtime changes have no effect and require a redeploy.