SequencePAR: Understanding pedestrian attributes via a sequence generation paradigm

Open Webpage

Jiandong Jin, Xiao Wang, Yin Lin, Chenglong Li, Lili Huang, Aihua Zheng, Jin Tang

Published: 01 Jan 2026, Last Modified: 14 Nov 2025Pattern Recognit. 2026EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•First work to handle PAR using sequence generation, addressing imbalanced and noisy attribute learning.•Autoregressive generation circumvents the issue of sparse attribute prediction due to many negative samples.•Novel masked Transformer decoder predicts each attribute sequentially using pedestrian tokens and textual representations, addressing weak attribute context in standard multi-label classification.