DualAgg: Dual-Branch Spatial Aggregation for Open-Vocabulary Semantic Segmentation

Le résumé fourni par la source

Open Vocabulary Semantic Segmentation (OVSS) aims to segment categories not fully observed during training, relying on effective vision–language alignment and semantic generalization. Existing methods typically apply early category pruning to reduce computational cost and use Transformer-based encoders for global context modeling, but suffer from two key limitations: static and coarse early rejection may prematurely discard relevant categories, and global aggregation tends to dilute fine-grained local structural information critical for accurate segmentation. To address these issues, we propose a novel framework integrating Adaptive Early Category Rejection (AECR) and Dual-Branch Spatial Aggregation (DBSA). AECR performs spatial-aware and uncertainty-guided progressive category filtering, while DBSA decouples global semantic aggregation from local structural modeling via parallel branches and adaptive fusion. Extensive experiments on multiple OVSS benchmarks demonstrate consistent improvements over strong baselines, particularly for rare and unseen categories.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.