Delineate Anything v2: A Global Foundation Model for Field Delineation
The model is built for national-scale field boundary mapping, with a 73-million-instance dataset behind it.
The paper says Delineate Anything v2 targets crop-field delineation where general vision models struggle with geospatial texture, topology, and scale. Its FBIS-73M dataset spans 61 countries, and the authors add a curation pipeline meant to handle merged administrative parcels and weak boundaries. They report a 0.284 [email protected] gain over prior state of the art and say Ukraine-wide mapping took 5.4 hours on a consumer workstation. Code, weights, data, and vector products are listed as publicly available. HF Daily Papers' note
The paper says Delineate Anything v2 targets crop-field delineation where general vision models struggle with geospatial texture, topology, and scale. Its FBIS-73M dataset spans 61 countries, and the authors add a curation pipeline meant to handle merged administrative parcels and weak boundaries. They report a 0.284 [email protected] gain over prior state of the art and say Ukraine-wide mapping took 5.4 hours on a consumer workstation. Code, weights, data, and vector products are listed as publicly available. HF Daily Papers' note
score 6