Segment Foreground (U2 Net)

How it fits

Typical backends

Caveats

• I/O contractThe output arm is a single downstream-resolved choice fixed at deploy time; every arm describes the same single foreground region with class id zero, so the mask, box and confidence stay consistent across arms.
• LimitationOnly one salient subject is segmented per frame; use a class- or prompt-driven segmenter when several distinct objects must be separated.
• Parameter interactionthreshold cuts the normalised saliency map into the binary mask; raise it to keep only the confident core and lower it to include soft edges.
• FallbackWhen nothing clears the cutoff the output still carries an all-false mask and a zero-area box, so frames are never dropped.
• AccuracySaliency is computed at a fixed small input size and resampled back to the source resolution, so very thin or hairline structures can be lost.
• State lifetimeThe bound ONNX checkpoint, device and threshold are captured once at startup; the `u2net.onnx` slot file is staged locally and loaded offline, so nothing is fetched from a remote at deploy or runtime.