GeoCueFormer: Geometry-Guided Wavelet Representation and Prediction-Cued Dual-Stage Decoder for Underwater Semantic Segmentation

Xian Wu; Xinjin Li; Yiliu Xu; Yining Liu; Yong Jiang

cs.CVSep 16, 2026

GeoCueFormer: Geometry-Guided Wavelet Representation and Prediction-Cued Dual-Stage Decoder for Underwater Semantic Segmentation

Authors: Xian Wu, Xinjin Li, Yiliu Xu, Yining Liu, Yong Jiang

Organizations: Southwest University of Science and Technology, Mianyang, China · Columbia University, New York, United States · Carnegie Mellon University, Pittsburgh, United States · University of California, Berkeley, CA, United States

Abstract

Underwater semantic segmentation is essential for marine ecosystem monitoring, yet remains challenging due to severe visual degradation. Light absorption and scattering often lead to color shifts, low contrast, and blurred boundaries, making shallow detail features unreliable. Existing underwater segmentation methods improve RGB feature aggregation or boundary prediction, but still lack an explicit mechanism to distinguish structure-related details from degradation-induced responses. To address this limitation, we propose GeoCueFormer, a lightweight framework that combines geometry-constrained frequency enhancement with prediction-cued refinement. GeoCueFormer performs stage-specific wavelet enhancement on hierarchical encoder features to complement shallow boundary details while preserving deep structural semantics. A depth-derived spatial gate constrains shallow frequency enhancement toward geometry-consistent regions, and a prediction-cued dual-stage decoder further refines ambiguous high-resolution features. GeoCueFormer obtains 82.23% and 73.04% mIoU on SUIM and DUT, respectively. Under comparable model complexity and standard benchmark settings on SUIM and DUT, it achieves SOTA performance while maintaining a favorable accuracy-complexity trade-off. These results show that distinguishing structural details from degradation-induced interference is more effective for underwater segmentation.

Figures & tables

Fig. 1: Motivation of depth-guided geometric cues.

Fig. 2: Overall architecture of GeoCueFormer.

Fig. 3: Stage-Specific Wavelet Representation with Geometry-Guided Gating

Fig. 4: Prediction-Cued Dual-Stage Decoder.

Method	GFLOPs $\downarrow$	Params $\downarrow$	mIoU (%) $\uparrow$
GeoCueFormer (Ours)	13.58	4.28M	82.23
UWSegFormer [ 9 ]	3.21	21.78M	82.12
LVT [ 10 ]	10.51	3.83M	80.72
Swin Transformer [ 3 ]	276	58.94M	80.70
SegFormer [ 6 ]	7.94	3.72M	80.57
MSNet [ 11 ]	N/R	6.2M	79.83

Table 1: Comparison with state-of-the-art methods on the SUIM dataset. N/R denotes unavailable results.

Method	GFLOPs ( $\downarrow$ )	Params ( $\downarrow$ )	mIoU (%)
GeoCueFormer (Ours)	13.58	4.28M	73.04
UWSegFormer [ 9 ]	3.21	21.78M	71.41
Lightweight DeepLabv3+ [ 1 ]	39.612	6.628M	71.18
SegFormer [ 6 ]	7.94	3.72M	70.81
UHRS-Net [ 8 ]	N/R	N/R	70.09
HRNetv2 [ 5 ]	90.972	29.540M	70.09

Table 2: Comparison with state-of-the-art methods on the DUT dataset. N/R denotes unavailable results.

Fig. 5: Qualitative comparison of segmentation results. Cyan boxes highlight challenging regions with low contrast, small objects, or ambiguous boundaries.

Variant	Wavelet	Depth	PCDD	mIoU (%)
SegFormer Baseline	$\times$	$\times$	$\times$	80.57
+ Wavelet	$\checkmark$	$\times$	$\times$	80.98
+ Two-stage Decoder	$\times$	$\times$	$\checkmark$	81.42
+ Wavelet + Two-stage Decoder	$\checkmark$	$\times$	$\checkmark$	81.21
+ Wavelet + Depth	$\checkmark$	$\checkmark$	$\times$	81.58
Full Model	$\checkmark$	$\checkmark$	$\checkmark$	82.23

Table 3: Ablation study of the main components on SUIM.

Variant	Frequency strategy	Decoder	mIoU (%)
Baseline	-	Single	80.57
Full-stage Full-band	LL/LH/HL/HH	Two-stage	80.91
GeoCueFormer	Specific / Selective	Two-stage	82.23

Table 4: Analysis of frequency-band selection strategy on SUIM.

Setting	Depth-guidance stages	mIoU (%)
GeoCueFormer	Stage 1 + Stage 2	82.23
Depth-All	Stage 1 + Stage 2 + Stage 3 + Stage 4	80.06

Table 5: Analysis of depth-guidance injection position on SUIM.

Setting	$\alpha_{\text{init}}$	mIoU (%)
GeoCueFormer	0.01	80.40
GeoCueFormer	0.05	81.44
GeoCueFormer	0.10	82.23
GeoCueFormer	0.20	81.09

Table 6: Analysis of the initial residual scale

\alpha_{\text{init}}

on SUIM.

Semantic Segmentation

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

See all 410 papers

GeoPrompt: Geometry-Conditioned Prompt Learning for Underwater Open-Vocabulary Instance Segmentation
Oct 25, 2026Shanshan Han, Yuxi Wang, Can Qin +4Instance Segmentation Prompt Learning
Ocean University of China, Qingdao, China · Nanjing University of Aeronautics and Astronautics, Nanjing, China
Learning Dynamic Structural Specialization for Underwater Salient Object Detection
May 15, 2026Lin Hong, Chenhui Wang, Linan Deng +7Underwater Imaging Salient Object Detection
School of Automation, Southeast University, Nanjing, China.
USIS-PGM: Photometric Gaussian Mixtures for Underwater Salient Instance Segmentation
Mar 14, 2026Lin Hong, Xiangtong Yao, Mürüvvet Bozkurt +2Image Segmentation Instance Segmentation
Department of Electronic and Computer Engineering, The Hong Kong University of Science and Technology, Hong Kong, China · School of CIT, Chair i6, Technical University of Munich, Munich, Germany · School of CIT, Technical University of Munich, Munich, Germany +1

Oct 25, 2026J/K move · Enter open · S save

GeoPrompt: Geometry-Conditioned Prompt Learning for Underwater Open-Vocabulary Instance Segmentation

Shanshan Han, Yuxi Wang, Can Qin, Yujie Chen+3

Ocean University of China, Qingdao, China · Nanjing University of Aeronautics and Astronautics, Nanjing, China

Instance Segmentation Prompt Learning Visual Prompt Learning Underwater Imaging

Open paper AskPDF arXiv

Open-vocabulary instance segmentation (OVIS) provides a promising solution for marine perception by recognizing and segmenting categories beyond predefined label spaces. Existing OVIS methods usually rely on vision-language alignment, where visual region features are matched with textual category embeddings to recognize both seen and unseen classes. However, this alignment assumption becomes fragile in underwater environments. Color attenuation, scattering, turbidity, and low contrast distort visual appearance, causing region features to deviate from their corresponding semantic concepts. As a result, the similarity between degraded visual features and textual classifiers becomes unreliable, which particularly weakens the recognition of unseen underwater categories that cannot be corrected by category-specific supervision. To address these issues, we propose GeoPrompt, a geometry-conditioned prompt learning framework for underwater open-vocabulary instance segmentation. Our key insight is that geometric structure is more stable than visual appearance in degraded underwater scenes. Specifically, GeoPrompt introduces a geometry-guided visual adaptation module (GVAM) to extract depth-derived structural priors and inject them into multi-scale visual features. It further develops a geometry-conditioned prompt learning (GCPL) that generates geometry-aware textual classifiers from structural representations. These geometry-aware classifiers are integrated with fixed and image-conditioned textual classifiers to improve open-vocabulary semantic alignment. Experiments on the MARIS-OpenIS benchmark show that GeoPrompt achieves state-of-the-art performance and brings consistent improvements on underwater open-vocabulary instance segmentation.

Abstract

Semantic Segmentation

Explore similar work

GeoPrompt: Geometry-Conditioned Prompt Learning for Underwater Open-Vocabulary Instance Segmentation

Learning Dynamic Structural Specialization for Underwater Salient Object Detection

USIS-PGM: Photometric Gaussian Mixtures for Underwater Salient Instance Segmentation