Reload-Mamba: Hierarchical Anti-Dilution State-Space Modeling for Multi-Class Semantic Segmentation

A new semantic segmentation framework called Reload-Mamba directly tackles a known weakness in state-space models for computer vision: the gradual loss of boundary and detail information during sequential processing. The approach, detailed in a paper submitted to arXiv on June 16, 2026, introduces a hierarchical anti-dilution mechanism to restore these critical signals for multi-class dense prediction tasks [1]. Mamba-based state space models are valued for their linear-time long-range modeling capabilities, making them suitable for high-resolution dense prediction. However, their sequential state-space propagation can attenuate boundary-sensitive and detail-sensitive responses, a problem the researchers term "response dilution" [2]. To counter this, Reload-Mamba incorporates three segmentation-specific designs. First, a boundary-supervised local detail prior is explicitly trained with ground-truth boundary masks to identify regions requiring response restoration. Second, a class-uncertainty-aware Reload Gate uses per-pixel class entropy from a pre-reload auxiliary head as an additional gating signal, a formulation the authors note is informative only under multi-class dense prediction. Third, a hierarchical multi-level Reload mechanism applies anti-dilution refinement at three decoder levels and fuses the restored representations top-down [2]. The framework is built upon a ConvNeXt-Tiny encoder with a multi-scale decoder and four-directional Mamba scanning with pixel-wise directional attention [2]. On the ADE20K benchmark, Reload-Mamba achieves 47.9% single-scale and 48.9% multi-scale mean Intersection over Union (mIoU). On the Cityscapes dataset, it reaches 83.2% single-scale mIoU. When configured with a ResNet-101 backbone and COCO pre-training under the standard DeepLab-style protocol, the model attains 87.8% mIoU on the PASCAL VOC 2012 validation set [2]. Controlled ablation studies underscore the contribution of each component. The researchers report that the three segmentation-specific designs cumulatively improve performance over a direct port of a prior anti-dilution architecture originally proposed for binarization by +2.2 mIoU on ADE20K [2]. The paper is available on arXiv, and the broader research community can interact with it through platforms such as Hugging Face, which allows users to link papers to models, datasets, and interactive demos, and even embed those demos directly on arXiv abstract pages [4][5].