Improved query specialization for transformer-based visual relationship detection

Download PDF
Open Webpage

Jongha Kim, Jihwan Park, Jinyoung Park, Jinyoung Kim, Sehyung Kim, Hyunwoo J. Kim

Published: 2026, Last Modified: 07 Feb 2026Inf. Sci. 2026EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•We propose SpeaQ+, a label assignment for training Transformer-based detectors.•SpeaQ+ provides specialized and abundant training signals for a detector.•SpeaQ+ improves seven baseline models across five benchmarks.