Almanac
other

vision-language grounding

otheractiveprovisionalvision-language-grounding-dedbbeed·1 events·first seen 21d ago

Aliases: vision-language grounding

Co-occurring entities

More like this (12)

Recent events (1)

6arXiv · cs.LG·21d ago·source ↗

LocateAnything: Parallel Box Decoding for Fast and Accurate Vision-Language Grounding

LocateAnything introduces Parallel Box Decoding (PBD), a method that decodes bounding boxes and points as atomic units in a single step rather than sequentially token-by-token, improving both throughput and geometric coherence in visual grounding tasks. The framework is paired with a large-scale data engine producing LocateAnything-Data, a 138-million-sample training dataset for high-precision localization. Evaluations show advances on the speed-accuracy frontier across diverse grounding and detection benchmarks. The work addresses a fundamental architectural mismatch in how current VLMs handle 2D spatial coordinates.