DroneSegEdit keeps the drone editor's mask but adds language verification, reaching 52.78 mIoU on RIS-LAD
Synopsis
In aerial drone imagery, a referred target often covers only tens of pixels and sits among visually similar vehicles or people, so a generative editor can draw a plausible mask around the wrong instance. DroneSegEdit freezes the open-weight LongCat-Image-Edit, trains only a lightweight LoRA to render candidate regions, then uses a language-grounded spatial verifier to strip foreground inside that candidate that the referring expression does not support. On the complete official RIS-LAD test set it reaches 52.78 mIoU, 7.16 points above the previous best ReSaP-8B, and stays competitive overall on UAVid-RIS and VDD-RIS, with the clearest gains on compact vehicles and humans.
Interpretation
Language-guided drone segmentation can now reuse a pretrained image editor's RGB output interface instead of building a separate discriminative segmentation decoder. Earlier referring segmentation relied on cross-modal decoders or on multimodal reasoning followed by a promptable segmentation model; applying a generative editor directly produced plausible masks that could land on the wrong instance among small, repeated objects. On the complete official RIS-LAD test set, DroneSegEdit reaches 52.78 mIoU and 54.01 oIoU, 5.42 and 6.92 points above the paired direct generations from the same editor and 7.16 and 0.57 points above ReSaP-8B.
The design splits drawing from choosing: the editor supplies candidate geometry, and language evidence decides only which foreground inside that candidate survives. The Language-Grounded Spatial Verifier (LGSV) reads four hierarchical hidden-state layers of the frozen Qwen2.5-VL backbone, pools the tokens of the referring expression into a condition, and FiLM-modulates the visual grid into a coarse language support map. Verification is constrained to the rendered candidate and only removes foreground; on RIS-LAD it raises precision from 52.93 to 62.10 while recall moves only from 81.04 to 80.57, and the predicted-to-ground-truth area ratio falls from 1.531 to 1.297.
The same RGB interface extends to full-scene parsing, rendering every class in one generation from a single task-level instruction. The dense VDD adapter receives neither per-class names nor palette colors; the class-to-color correspondence is learned from paired RGB targets during LoRA fine-tuning. On the official 280/80/40 split, this verifier-free version reaches 80.96 mIoU, second only to UPerNet–Swin-L at 82.17 under the matched-resolution comparison.
Perspective
Experiments cover the three referring benchmarks RIS-LAD, UAVid-RIS and VDD-RIS plus dense VDD parsing, each with an independently trained adapter under its official training and evaluation protocol, with checkpoints and decision thresholds selected on validation only; the intended readers are researchers in drone vision and generative segmentation.
Verification is proposal-constrained, so it cannot recover targets the editor omitted and cannot reliably separate candidates whose evidence is indistinguishable; on VDD-RIS its oIoU trails AeroReformer by 3.80 points and P@0.9 drops from 44.65 to 41.33 relative to the raw editor, and the authors list color collisions, coarse evidence and reduced-resolution dense evaluation as limitations.