CRISP: Calibration-Aware Visual State Space Duality for Remote Sensing Semantic Segmentation

Kangning Wang; Haopeng Zhang; Zhiguo Jiang

cs.CVAug 24, 2026

CRISP: Calibration-Aware Visual State Space Duality for Remote Sensing Semantic Segmentation

Authors: Kangning Wang, Haopeng Zhang, Zhiguo Jiang

Organizations: Tianmushan Laboratory, Beihang University, Hangzhou 311115, China · Hangzhou International Innovation Institute, Beihang University, Hangzhou 311115, China · School of Astronautics, Beihang University, Beijing 102206, China · Key Laboratory of Spacecraft Design Optimization and Dynamic Simulation Technology, Ministry of Education, Beijing 102206, China

Abstract

State space models, especially Visual State Space Duality (VSSD), have emerged as efficient linear-time alternatives to Transformers for dense visual tasks. However, we observe that VSSD compresses spatial context into a global aggregation that suppresses high-frequency responses, causing excessive boundary smoothing in remote sensing semantic segmentation. To address this, we propose CRISP, a calibration framework with two components. Its core, the Duality Calibration Operator (DCO), restores local contrast and boundary responses through residual injection and frequency calibration within the VSSD backbone, without altering its linear complexity. To retain the recovered detail, an Orthogonal Multi-Prototype (OMP) head assigns multiple orthogonally constrained prototypes per class to model large intra-class variance. Extensive experiments on Potsdam, Vaihingen, and LoveDA show that, with approximately 30M parameters, CRISP achieves consistent gains in mean F1 (mF) and mIoU while remaining competitive with state-of-the-art methods. Code is available at https://github.com/crazylifeha/CRISP.

Figures & tables

Figure 2: Overall CRISP architecture. DCO calibrates high-frequency responses inside the VSSD backbone, while OMP performs dynamic token generation and multi-prototype matching.

Method	Params (M)	Bg	Bu	Rd	Wa	Ba	Fo	Ag	mIoU
ABCNet [ 22 ]	13.6	49.78	46.56	48.08	63.83	20.03	38.33	47.04	44.81
VSSD [ 34 ]	56.1	50.41	50.96	51.86	64.94	24.00	40.57	47.35	47.15
DCSwin [ 40 ]	45.6	49.70	56.95	47.92	66.14	27.53	42.04	43.56	47.69
PyramidMamba [ 39 ]	125.1	50.73	51.06	51.47	67.81	26.78	40.08	49.17	48.16
BANet [ 38 ]	12.7	53.10	59.52	48.83	65.06	23.09	41.12	49.00	48.53
SFA-Net [ 18 ]	10.7	52.58	58.73	53.53	66.84	27.31	39.16	48.85	49.57

Table 1: Quantitative comparison on LoveDA. Class-wise IoU (%) and mIoU are reported; best and second-best results are bolded and underlined .

Method	Params (M)	Imp.	Build.	Low-veg.	Tree	Car	mFscore	mIoU
BANet	12.7	92.31	94.42	86.27	84.93	94.95	90.58	83.04
PyramidMamba	125.1	93.41	95.93	86.77	85.86	95.54	91.50	84.62
DCSwin	45.6	93.57	96.13	87.51	87.03	96.12	92.07	85.56
VSSD	56.1	93.78	96.27	87.13	87.20	95.97	92.07	85.57
SFA-Net	10.7	94.06	96.63	87.96	87.28	96.46	92.48	86.28
ABCNet	13.6	94.14	96.88	88.07	88.04	96.13	92.65	86.55

Table 2: Quantitative comparison on ISPRS Potsdam. Best/second-best results are bolded / underlined .

Method	Params (M)	Imp.	Build.	Low-veg.	Tree	Car	mFscore	mIoU
SFA-Net	10.7	91.76	94.35	82.42	88.43	75.73	86.54	76.87
UNetFormer	11.7	92.23	94.65	83.75	89.04	79.16	87.77	78.64
DCSwin	45.6	91.85	94.30	83.96	89.18	80.68	87.99	78.92
PyramidMamba	125.1	92.38	94.99	83.49	88.88	81.41	88.23	79.32
BANet	12.7	91.79	94.41	83.21	89.03	83.70	88.43	79.54
VSSD	56.1	92.40	94.94	82.86	88.54	83.96	88.54	79.75

Table 3: Quantitative comparison on ISPRS Vaihingen. Best/second-best results are bolded / underlined .

Config.	Res. Inj.	Hi-Pass	DC/HC	mIoU
Baseline				87.32
Res. Inj. only	✓			87.65
Hi-Pass only		✓		87.68
DC/HC only			✓	87.64
Hi-Pass + DC/HC		✓	✓	87.83
Full DCO	✓	✓	✓	88.10

Table 4: DCO sub-component ablation on Potsdam (UPerHead, single-scale).

Dec.	DCO	Orth	Margin	mIoU
UPer				87.32
UPer	✓			88.10
OMP		✓	✓	85.63
OMP	✓			86.85
OMP	✓	✓		88.06
OMP	✓		✓	88.02

Table 5: End-to-end component ablation on Potsdam (VSSD backbone).

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Param. (M)	FLOPs (G)	Mem. (G)	Lat. (ms)	FPS
SFA-Net	10.70	7.22	0.13	6.36	157.2
UNetFormer	11.72	11.76	0.10	2.88	347.7
BANet	12.71	15.09	0.48	4.35	230.1
ABCNet	13.63	15.59	0.47	2.55	392.1
AerialFormer	43.16	49.94	0.30	6.95	144.0
DCSwin	45.63	48.41	0.65	7.26	137.8

Appendix

Table S1 : Efficiency on Potsdam with

512\times 512

input.

Method / Metric	Encoder	DCO	Decoder	Total
VSSD baseline FLOPs (G)	25.07	–	210.58	235.65
CRISP Params (M)	26.20	$\!<\!10^{-3}$	6.13	32.32
CRISP FLOPs (G)	45.83	3.19	15.78	64.80

Appendix

Table S2 : Component-wise parameter and FLOPs breakdown.

Figure S1 : Error maps comparing VSSD and CRISP. Red pixels indicate prediction errors.

Figure S2 : Feature activation maps before and after DCO.

$K$	mIoU (%)	mFscore (%)
2	88.10	93.56
3	88.25	93.64
4	88.17	93.59
5	88.19	93.60

Appendix

Table S3 : Effect of prototype number

K

on Potsdam.

Model	Band mIoU	Band Acc	mBF
w/o DCO	0.6830	0.7833	0.8698
w/o proto. regs	0.6848	0.7840	0.8232
Full CRISP	0.7179	0.8104	0.8953

Appendix

Table S4 : Boundary-band diagnostic on Potsdam.

Dataset	VSSD	CRISP	Gain
Potsdam	0.6669	0.7038	+0.0369
Vaihingen	0.6730	0.6935	+0.0205
LoveDA	0.1999	0.2239	+0.0240

Appendix

Table S5 : Cross-dataset class-aware mBF (CRISP vs. VSSD baseline).

Figure S3 : Detailed VSSD core with DCO, complementing the compact main-paper diagram.

Semantic Segmentation

Semantic segmentation assigns a class label to every image pixel, enabling detailed scene understanding across domains and sensors.

See all 414 papers

Efficient Remote Sensing Instance Segmentation with Linear-Time State Space Distilled Visual Foundation Models
Jun 24, 2026Qinzhe Yang, Keyan Chen, Jia Xu +2Remote Sensing Image Segmentation Instance Segmentation
Shen Yuan Honors College, Beihang University, Beijing 100191, China · Department of Aerospace Intelligent Science and Technology, School of Astronautics, and with the State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, Beijing 100191, China · Qian Xuesen Laboratory of Space Technology, China Academy of Space Technology, Beijing 100094, China
Decouple and Rectify: Semantics-Preserving Structural Enhancement for Open-Vocabulary Remote Sensing Segmentation
Apr 2, 2026Jie Feng, Fengze Li, Junpeng Zhang +4Remote Sensing Image Segmentation Open-Vocabulary Semantic Segmentation
Xidian University China · Jimei University China · South China University of Technology China
Open-Vocabulary Semantic Segmentation in Remote Sensing via Hierarchical Attention Masking and Model Composition
Feb 27, 2026Mohammadreza Heidarianbaei, Mareike Dorozynski, Hubert Kanyamahanga +2Open-Vocabulary Semantic Segmentation Remote Sensing Image Segmentation
Institute of Photogrammetry and GeoInformation Leibniz University Hannover Germany

Jun 24, 2026J/K move · Enter open · S save

Efficient Remote Sensing Instance Segmentation with Linear-Time State Space Distilled Visual Foundation Models

Qinzhe Yang, Keyan Chen, Jia Xu, Zhenwei Shi+1

Shen Yuan Honors College, Beihang University, Beijing 100191, China · Department of Aerospace Intelligent Science and Technology, School of Astronautics, and with the State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, Beijing 100191, China · Qian Xuesen Laboratory of Space Technology, China Academy of Space Technology, Beijing 100094, China

Remote Sensing Image Segmentation Instance Segmentation Remote Sensing Knowledge Distillation State Space Models

Open paper AskPDF arXiv

The computational complexity of Transformers scales quadratically with the number of tokens, which significantly constrains the efficiency of vision models, particularly recent ViT-based foundation models in dense prediction tasks. Instance segmentation, a typical dense visual prediction task in the remote sensing field, faces similar challenges. In this paper, inspired by the recent advances of knowledge distillation in large language models, we introduce RS4D - a new remote sensing instance segmentation method with linear computational complexity, which addresses the inefficiency of long sequence modeling through distilled state space modeling (SSM). We propose an adaptive noise and masking knowledge distillation training method for pre-training lightweight SSM backbones, which effectively compresses knowledge from the vast self-attention space into a compact, dense linear state space. We also design a remote sensing image instance segmentation architecture based on this lightweight visual encoder, where we explore variants of three different backbones and two segmentation heads. Extensive experiments are conducted on multiple benchmark datasets, including SSDD, WHU, and NWPU. Compared to ViT-based approaches, our proposed SSM backbone achieves an 8x reduction in parameters and a 9x reduction in FLOPs while maintaining comparable or superior accuracy to both ViT- and CNN-based instance segmentation methods. The implementation codes have been publicly available at https://github.com/QinzheYang/RS4D.

Abstract

Semantic Segmentation

Explore similar work

Efficient Remote Sensing Instance Segmentation with Linear-Time State Space Distilled Visual Foundation Models

Decouple and Rectify: Semantics-Preserving Structural Enhancement for Open-Vocabulary Remote Sensing Segmentation

Open-Vocabulary Semantic Segmentation in Remote Sensing via Hierarchical Attention Masking and Model Composition