Instead of squishing the full frame to a square, optionally crops a square
region around the mask bounding box (union across all frames) before resizing.
Preserves aspect ratio of the subject and gives the model a focused,
undistorted view.
New optional inputs on SelVA Feature Extractor:
- crop_to_mask (BOOLEAN, default false) — enable the crop path
- crop_margin (FLOAT 0–1, default 0.1) — padding around the bbox as a
fraction of the bounding box side
_compute_mask_bbox: resizes mask to frame resolution, takes union over
all mask frames, expands to square + margin, shifts into frame bounds to
preserve square shape before clamping. Falls back to center square crop
if mask is empty.
Bbox is computed once from the original-resolution mask and reused for
both CLIP (384px) and sync (224px) frame sequences. Combine with
mask_clip/mask_sync for full background suppression on top of the crop.
Cache hash includes crop_to_mask and crop_margin when mask is connected.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Ethanfel
f3cabcad90