FrothSeg — Flotation-Froth Instance Segmentation Lab (Browser ML)
Published:
A browser-native lab for instance segmentation of flotation froth: delineating individual bubbles in a froth image, entirely client-side via ONNX. Live at frothseg.fasl-work.com, part of the Faena hub.
A real model, and a real classical ladder
Two tiers on the same cases: seven classical delineators (watershed and related morphology), and LamellaStar, a genuine four-head research model shipped as a three-seed logit-mean ensemble (top config N1: mean AP 0.5186, AP50 0.8279, panoptic quality 0.7359), with SAM2 and Cellpose-SAM as offline teachers (checkpoints checksum-recorded) distilled to a compact mask head for the browser.
The honest core: the data does not exist, and it says so
The field’s real blocker is that there is no public labelled real froth dataset. FrothSeg states this openly, records the data search as a null (Roboflow dropped), keeps the froth cases synthetic and labelled as such, and adopts BBBC038 (64 real dense-touching images, CC0) as an adjacent-domain transfer probe, explicitly noting that an adjacent real domain does not clear the froth blocker. The transfer study even refuted its own hypothesis: the synthetic froth ranking is generator-specific and the model drops on real transfer. That published null, not a flattering synthetic score, is the deliverable. A research lab, not a plant tool.