📝 SummarXiv

Abstract

Recently, Referring Image Segmentation (RIS) frameworks that pair the Multimodal Large Language Model (MLLM) with the Segment Anything Model (SAM) have achieved impressive results. However, adapting MLLM to segmentation is computationally intensive, primarily due to visual token redundancy. We observe that traditional patch-wise visual projectors struggle to strike a balance between reducing the number of visual tokens and preserving semantic clarity, often retaining overly long token sequences to avoid performance drops. Inspired by text tokenizers, we propose a novel semantic visual projector that leverages semantic superpixels generated by SAM to identify "visual words" in an image. By compressing and projecting semantic superpixels as visual tokens, our approach adaptively shortens the token sequence according to scene complexity while minimizing semantic loss in compression. To mitigate loss of information, we propose a semantic superpixel positional embedding to strengthen MLLM's awareness of superpixel geometry and position, alongside a semantic superpixel aggregator to preserve both fine-grained details inside superpixels and global context outside. Experiments show that our method cuts visual tokens by 93% without compromising performance, notably speeding up MLLM training and inference, and outperforming existing compressive visual projectors on RIS.