DGSeg: Dynamic Gating of Semantic-Spatial Guided Predictions for Reasoning Segmentation

2026-07-06

AI Summary

Existing reasoning-based segmentation methods rely on sparse intermediate cues—such as points or bounding boxes—which are prone to ambiguity and noise, making it difficult to accurately align with the nuanced intent of complex language queries. To address this limitation, this work proposes DGSeg, a novel framework that decouples semantic and spatial cues and leverages a multimodal large language model to generate complementary guidance signals for a dual-branch segmentation network. A lightweight dynamic gating module adaptively fuses the branch predictions based on their relative quality under supervision. The proposed method achieves gIoU scores of 69.6% and 67.3% on the ReasonSeg validation and test sets, respectively, significantly outperforming strong baselines.

Abstract

Reasoning segmentation aims to predict pixel-wise masks for targets given complex language queries. Existing approaches leverage Multimodal Large Language Models (MLLMs) for vision-language reasoning and generate intermediate target cues (e.g., points or boxes) to guide a segmentation model. However, compressing rich reasoning into sparse cues often introduces ambiguity and noise, preventing these cues from accurately preserving the reasoning intent. While multiple complementary cues can enrich target information, existing methods typically feed them jointly into a single segmentation process, allowing ambiguous or erroneous cues to affect the entire prediction. Therefore, we propose DGSeg, a reasoning segmentation framework that learns to fuse predictions guided by semantic and spatial cues. Specifically, the MLLM jointly reasons about both target identity and spatial location, producing complementary semantic and spatial cues that are fed into separate segmentation branches. Their predictions are adaptively integrated by a lightweight dynamic gating module trained with relative branch-quality supervision to suppress noisy or conflicting regions. Extensive experiments demonstrate that DGSeg consistently outperforms strong baselines on multiple benchmarks and achieves 69.6% and 67.3% gIoU on the challenging ReasonSeg validation and test splits. Code is available at https://github.com/RZZeng/DGSeg.

Problem

• reasoning segmentation
• multimodal large language models
• semantic-spatial cues
• ambiguous cues
• segmentation guidance

Innovation

• Dynamic Gating
• Reasoning Segmentation
• Semantic-Spatial Cues
• Multimodal Large Language Models
• Adaptive Fusion