H-SPAM: Hierarchical Superpixel Anything Model

Julien Walther; Rémi Giraud; Michaël Clément

cs.CVApr 13, 2026

H-SPAM: Hierarchical Superpixel Anything Model

Authors: Julien Walther, Rémi Giraud, Michaël Clément

Organizations: Univ. Bordeaux, CNRS, Bordeaux INP, IMS, UMR 5218, France · Univ. Bordeaux, CNRS, Bordeaux INP, LaBRI, UMR 5800, France

Abstract

Superpixels offer a compact image representation by grouping pixels into coherent regions. Recent methods have reached a plateau in terms of segmentation accuracy by generating noisy superpixel shapes. Moreover, most existing approaches produce a single fixed-scale partition that limits their use in vision pipelines that would benefit multi-scale representations. In this work, we introduce H-SPAM (Hierarchical Superpixel Anything Model), a unified framework for generating accurate, regular, and perfectly nested hierarchical superpixels. Starting from a fine partition, guided by deep features and external object priors, H-SPAM constructs the hierarchy through a two-phase region merging process that first preserves object consistency and then allows controlled inter-object grouping. The hierarchy can also be modulated using visual attention maps or user input to preserve important regions longer in the hierarchy. Experiments on standard benchmarks show that H-SPAM strongly outperforms existing hierarchical methods in both accuracy and regularity, while performing on par with most recent state-of-the-art non-hierarchical methods. Code and pretrained models are available: https://github.com/waldo-j/hspam.

Figures & tables

Panel of Figure 1 — Figure 1: Hierarchical superpixel segmentation example. The proposed H-SPAM method generates very regular and easily interpretable superpixels that are aligned with the image objects through the hierarchy, compared to HHTS [ 8 ] .

Figure 2: Global framework of the H-SPAM method. An object-based superpixel segmentation provides a fine superpixel map along with a high-level object map and features that are used by to create a hierarchy that respects the image objects through the merging process. Visual attention can also be used to concentrate or delay the merge on specific areas.

Panel of Figure 3 — Figure 3: Object-based superpixel region adjacency graph . (b) Superposition of the prior object map, the initial superpixels, and the edges between the superpixels. The red ( – ) and green ( – ) connections respectively mean that the superpixels do not belong or belong to the same object.

Figure 4: Object-based hierarchy creation . Our merging process combines 2 distinct phases corresponding to the processing of high-level prior objects. The first one only allow merges inside objects while the second one merges objects.

Panel of Figure 5 — Figure 5: Illustration of the attention modes . With our object-based framework (middle row), the attention can be averaged within objects to provide a cleaner guide for the merging process. The bottom row shows the user interactive mode, where red/white crosses lead to fewer/more superpixels in an object, offering a complementary alternative to our multi-scale segmentation method.

Figure 6: Ablation study - Influence of parameters. (a) Importance of the prior mask. Using object priors to guide the hierarchy largely improves the segmentation accuracy. (b) Influence of

w_{pos}

on the accuracy and regularity.

Panel of Figure 7 — Figure 7: Influence of the spatial parameter $w_{pos}$ for K=500 superpixels. This parameter controls the shape consistency of the decomposition. A high $w_{pos}$ enforces smallest shapes to merge. By default, we use $w_{pos}=5$ .

Figure 8: Quantitative comparison of H-SPAM to state-of-the-art hierarchical superpixel methods. H-SPAM is the most accurate while being also the most regular hierarchical methods on the three datasets.

Figure 9: Quantitative comparison of H-SPAM to state-of-the-art non-hierarchical superpixel methods on the BSD. Although constrained to be hierarchical, H-SPAM is one of the best performing method in both accuracy and regularity, preserving the performance of single scale object-based approach [ 29 ] .

Panel of Figure 10 — Figure 10: Qualitative example of H-SPAM for different scales. From left to right: 1250, 500, 150, 50 superpixels. H-SPAM produces perfectly nested, very regular and easily interpretable regions, that align well with object boundaries.

Panel of Figure 11 — Figure 11: Qualitative comparison between hierarchical methods. Number of superpixels from left to right: 1250, 800, 500, 150, 50. H-SPAM produces regular and easily interpretable regions, that align well with object boundaries compared to other methods.

Figure 12: Influence of the number of superpixel: Lower

N_{f}

performs better at small

K

, while higher

N_{f}

becomes better at intermediate

K

Panel of Figure 13 — Figure 13: Illustration of the attention modes . With our object-based framework (middle row), the attention can be averaged within objects to provide a cleaner guide for the merging process. The bottom row shows the user interactive mode, where red/white crosses lead to fewer/more superpixels in the object.

Panel of Figure 14 — Figure 14: Qualitative comparison between hierarchical methods. Number of superpixels from left to right: 1250, 800, 500, 150, 50. H-SPAM produces regular and easily interpretable regions, that align well with object boundaries compared to other methods.

Panel of Figure 15 — Figure 15: Qualitative comparison between hierarchical methods. Number of superpixels from left to right: 1250, 800, 500, 150, 50. H-SPAM produces regular and easily interpretable regions, that align well with object boundaries compared to other methods..

Panel of Figure 16 — Figure 16: Qualitative example of H-SPAM for different scales. From left to right: 1250, 500, 150, 50 superpixels. H-SPAM produces perfectly nested, very regular and easily interpretable regions, that align well with object boundaries compared to other methods.

Image Segmentation

Image segmentation assigns pixels or voxels to objects and regions, supporting measurement, scene understanding, and decisions across image domains.

See all 390 papers

Square Superpixel Generation and Representation Learning via Granular Ball Computing
Mar 31, 2026Shuyin Xia, Meng Yang, Dawei Dai +6Visual Representation Learning
Shuyin Xia, Meng Yang, Dawei Dai, Fan Chen, Shilin Zhao and Junwei Han are with Chongqing University of Posts and Telecommunications, Chongqing, China. · Xinbo Gao and Wen Lu are with Xidian University, Xi’an, China. · Guoyin Wang is with Chongqing Normal University, Chongqing, China.
SPARC: SuperPixel-Aware Region Contrastive Learning for Self-Supervised Dense Prediction
Sep 16, 2026David Szczecina, Yuanpei Xiang, Jitao Hu +4Self-Supervised Visual Representation Learning Contrastive Learning
Systems Design Engineering University of Waterloo Waterloo, Ontario, Canada
SP-MoMamba: Superpixel-driven Mixture of State Space Experts for Efficient Image Super-Resolution
May 25, 2026Wenbin Zou, Yawen Cui, Yi Wang +5Visual SSMs Mixture-of-Experts Models
Department of Electrical and Electronic Engineering, The Hong Kong Polytechnic University, Hong Kong SAR · School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China

Mar 31, 2026J/K move · Enter open · S save

Square Superpixel Generation and Representation Learning via Granular Ball Computing

Shuyin Xia, Meng Yang, Dawei Dai, Fan Chen+5

Shuyin Xia, Meng Yang, Dawei Dai, Fan Chen, Shilin Zhao and Junwei Han are with Chongqing University of Posts and Telecommunications, Chongqing, China. · Xinbo Gao and Wen Lu are with Xidian University, Xi’an, China. · Guoyin Wang is with Chongqing Normal University, Chongqing, China.

Visual Representation Learning

Open paper AskPDF arXiv

Superpixels provide a compact region-based representation that preserves object boundaries and local structures, and have therefore been widely used in a variety of vision tasks to reduce computational cost. However, most existing superpixel algorithms produce irregularly shaped regions, which are not well aligned with regular operators such as convolutions. Consequently, superpixels are often treated as an offline preprocessing step, limiting parallel implementation and hindering end-to-end optimization within deep learning pipelines. Motivated by the adaptive representation and coverage property of granular-ball computing, we develop a square superpixel generation approach. Specifically, we approximate superpixels using multi-scale square blocks to avoid the computational and implementation difficulties induced by irregular shapes, enabling efficient parallel processing and learnable feature extraction. For each block, a purity score is computed based on pixel-intensity similarity, and high-quality blocks are selected accordingly. The resulting square superpixels can be readily integrated as graph nodes in graph neural networks (GNNs) or as tokens in Vision Transformers (ViTs), facilitating multi-scale information aggregation and structured visual representation. Experimental results on downstream tasks demonstrate consistent performance improvements, validating the effectiveness of the proposed method.

Abstract

Image Segmentation

Explore similar work

Square Superpixel Generation and Representation Learning via Granular Ball Computing

SPARC: SuperPixel-Aware Region Contrastive Learning for Self-Supervised Dense Prediction

SP-MoMamba: Superpixel-driven Mixture of State Space Experts for Efficient Image Super-Resolution